Studies in Computational Intelligence 127

Régis Gras · Einoshin Suzuki

Fabrice Guillet · Filippo Spagnolo
(Eds.)

Statistical
Implicative
Analysis
Theory and Applications

1฀3
Régis Gras, Einoshin Suzuki, Fabrice Guillet and Filippo Spagnolo (Eds.)
Statistical Implicative Analysis
Studies in Computational Intelligence, Volume 127
Editor-in-chief
Prof. Janusz Kacprzyk
Systems Research Institute
Polish Academy of Sciences
ul. Newelska 6
01-447 Warsaw
Poland
E-mail: [email protected]

Further volumes of this series can be found on our Vol. 117. Da Ruan, Frank Hardeman and Klaas van der Meer
homepage: springer.com (Eds.)
Intelligent Decision and Policy Making Support Systems,
Vol. 107. Margarita Sordo, Sachin Vaidya and Lakhmi C. Jain 2008
(Eds.) ISBN 978-3-540-78306-0
Advanced Computational Intelligence Paradigms Vol. 118. Tsau Young Lin, Ying Xie, Anita Wasilewska
in Healthcare - 3, 2008 and Churn-Jung Liau (Eds.)
ISBN 978-3-540-77661-1 Data Mining: Foundations and Practice, 2008
Vol. 108. Vito Trianni ISBN 978-3-540-78487-6
Evolutionary Swarm Robotics, 2008 Vol. 119. Slawomir Wiak, Andrzej Krawczyk and Ivo Dolezel
ISBN 978-3-540-77611-6 (Eds.)
Vol. 109. Panagiotis Chountas, Ilias Petrounias and Janusz Intelligent Computer Techniques in Applied
Kacprzyk (Eds.) Electromagnetics, 2008
Intelligent Techniques and Tools for Novel System ISBN 978-3-540-78489-0
Architectures, 2008 Vol. 120. George A. Tsihrintzis and Lakhmi C. Jain (Eds.)
ISBN 978-3-540-77621-5 Multimedia Interactive Services in Intelligent Environments,
2008
Vol. 110. Makoto Yokoo, Takayuki Ito, Minjie Zhang,
ISBN 978-3-540-78491-3
Juhnyoung Lee and Tokuro Matsuo (Eds.)
Electronic Commerce, 2008 Vol. 121. Nadia Nedjah, Leandro dos Santos Coelho
ISBN 978-3-540-77808-0 and Luiza de Macedo Mourelle (Eds.)
Quantum Inspired Intelligent Systems, 2008
Vol. 111. David Elmakias (Ed.)
ISBN 978-3-540-78531-6
New Computational Methods in Power System Reliability,
2008 Vol. 122. Tomasz G. Smolinski, Mariofanna G. Milanova
ISBN 978-3-540-77810-3 and Aboul-Ella Hassanien (Eds.)
Applications of Computational Intelligence in Biology, 2008
Vol. 112. Edgar N. Sanchez, Alma Y. Alanı́s and Alexander ISBN 978-3-540-78533-0
G. Loukianov
Discrete-Time High Order Neural Control: Trained with Vol. 123. Shuichi Iwata, Yukio Ohsawa, Shusaku Tsumoto,
Kalman Filtering, 2008 Ning Zhong, Yong Shi and Lorenzo Magnani (Eds.)
ISBN 978-3-540-78288-9 Communications and Discoveries from Multidisciplinary
Data, 2008
Vol. 113. Gemma Bel-Enguix, M. Dolores Jiménez-López ISBN 978-3-540-78732-7
and Carlos Martı́n-Vide (Eds.)
New Developments in Formal Languages and Applications, Vol. 124. Ricardo Zavala Yoe
2008 Modelling and Control of Dynamical Systems: Numerical
ISBN 978-3-540-78290-2 Implementation in a Behavioral Framework, 2008
ISBN 978-3-540-78734-1
Vol. 114. Christian Blum, Maria José Blesa Aguilera, Andrea
Roli and Michael Sampels (Eds.) Vol. 125. Larry Bull, Ester Bernadó-Mansilla
Hybrid Metaheuristics, 2008 and John Holmes (Eds.)
ISBN 978-3-540-78294-0 Learning Classifier Systems in Data Mining, 2008
ISBN 978-3-540-78978-9
Vol. 115. John Fulcher and Lakhmi C. Jain (Eds.)
Computational Intelligence: A Compendium, 2008 Vol. 126. Oleg Okun and Giorgio Valentini (Eds.)
ISBN 978-3-540-78292-6 Supervised and Unsupervised Ensemble Methods and their
Applications, 2008
Vol. 116. Ying Liu, Aixin Sun, Han Tong Loh, Wen Feng Lu ISBN 978-3-540-78980-2
and Ee-Peng Lim (Eds.)
Advances of Computational Intelligence in Industrial Vol. 127. Régis Gras, Einoshin Suzuki, Fabrice Guillet
Systems, 2008 and Filippo Spagnolo (Eds.)
ISBN 978-3-540-78296-4 Statistical Implicative Analysis, 2008
ISBN 978-3-540-78982-6
Régis Gras
Einoshin Suzuki
Fabrice Guillet
Filippo Spagnolo
(Eds.)

Statistical Implicative Analysis

Theory and Applications

With 147 Figures and 74 Tables

123
Régis Gras Einoshin Suzuki
LINA, FRE 2729 CNRS Department of Informatics
14 avenue de la Chaise Kyushu University
35170 Bruz, France 744 Motooka, Nishi, Fukuoka
[email protected] 819-0395, Japan
[email protected]

Fabrice Guillet Filippo Spagnolo

LINA, FRE 2729 CNRS Dipartimento di Matematica
Polytech’Nantes Univesrità di Palermo
rue C. pauc, BP 50609 Via Archirafi n.34
44306 Nantes, cedex 3, France 90123 Palermo, Italy
[email protected] [email protected]

ISBN 978-3-540-78982-6 e-ISBN 978-3-540-78983-3

Studies in Computational Intelligence ISSN 1860-949X

Library of Congress Control Number: 2008924359

c 2008 Springer-Verlag Berlin Heidelberg

°

This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broad-
casting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of
this publication or parts thereof is permitted only under the provisions of the German Copyright Law
of September 9, 1965, in its current version, and permission for use must always be obtained from
Springer-Verlag. Violations are liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
Cover design: Deblik, Berlin, Germany

Printed on acid-free paper

9 8 7 6 5 4 3 2 1

springer.com
Preface

Statistical implicative analysis is a data analysis method created by Régis Gras

almost thirty years ago which has a significant impact on a variety of areas
ranging from pedagogical and psychological research to data mining. This
new concept has developed into a unifying methodology, and has generated a
powerful convergence of thought between mathematicians, statisticians, psy-
chologists, specialists in pedagogy and last, but not least, computer scientists
specialized in data mining.
Statistical implicative analysis (SIA) provides a framework for evaluating
the strength of implications; such implications are formed through common
knowledge acquisition techniques in any learning process, human or artificial.
Therefore, the epistemological interest of SIA is, in my opinion, of universal
interest for researchers. In many applications implications appear as “rules”
and, as it is often the case, rules have exceptions. SIA provides a powerful
instrument for quantifying the quality of a rule taking into account the real-
ity of these exceptions. Many applications, especially in data mining, extract
large sets of rules that are impossible to assimilate by humans and used effi-
ciently in decision processes. Therefore, it is important to develop measures
of interestingness for these rules and the success of SIA-based techniques in
this direction is indisputable.
This volume collects significant research contributions of several rather
distinct disciplines that benefit from SIA. Contributions range from psycho-
logical and pedagogical research, bioinformatics, knowledge management, and
data mining.
The first applications of SIA were in the realm of didactics and this field
is richly represented here by several contributions that focus on such diverse
problems as didactics of algebra and geometry, the teaching of functions rep-
resentations and graphing, Bayesian inference, and student representations of
physical activities.
Interesting data mining applications authored by leading researchers in the
field range from applying SIA in the study of rules produced by decision trees,
association rules generated by the analysis of transactional data, temporal
VI Preface

rules, measures of interestingness for various types of rules, and hierarchical

organization of rules. A novel method for analyzing DNA microarrays is for-
mulated using SIA concepts. Furthermore, applications of SIA to the study of
ontologies and textual taxonomies, as well as applications to fuzzy knowledge
discovery are also included.
We have here a new volume that confirms the validity of a novel and
powerful statistical methodology, though many convincing applications. The
contributors have done a masterful job of exposition.
After reading this book, I have in mind a few applications of SIA in my own
research. I am convinced that the readers will find this volume as stimulating
as I did.

Boston, Prof. Dan A. Simovici

September, 2007 Department of Computer Science
University of Massachusetts at Boston University
 Preface VII

Review Committee
All published chapters have been reviewed by at least 2 referees.
• Saddo Ag Almouloud (University of Sao Paulo, Brazil)
• Carmen Batanero (University of Grenada)
• Hans Bock (Aachen University, Germany)
• Henri Briand (LINA, University of Nantes, France)
• Guy Brousseau (University of Bordeaux 3, France)
• Alex Freitas (University of Kent, UK)
• Athanasios Gagatsis (University of Chyprius)
• Robin Gras (University of Windsor, Canada)
• Howard Hamilton (University of Regina, Canada)
• Jiawei Han (University of Illinois, USA)
• David J. Hand (Imperial College, London, UK)
• André Hardy (University of Namur, Belgium)
• Robert Hilderman (University of Regina, Canada)
• Yves Kodratoff (LRI, University of Paris-Sud, France)
• Pascale Kuntz (LINA, University of Nantes, France)
• Ludovic Lebart (ENST, Paris, France)
• Amédéo Napoli (LORIA, University of Nancy, France)
• Maria-Gabriella Ottaviani (University of Roma, Italy)
• Balaji Padmanabhan (University of Pennsylvania, USA)
• Jean-Paul Rasson (University of Namur, Belgium)
• Jean-Claude Régnier (University of Lyon 2, France)
• Gilbert Ritschard (Geneve University, Switzerland)
• Lorenza Saitta (University of Piemont, Italy)
• Gilbert Saporta (CNAM, Paris, France)
• Dan Simovici (University of Massachusetts Boston, USA)
• Djamel Zighed (ERIC, University of Lyon 2, France)

Associated Reviewers
Nadja Maria Carmen Diaz, Philippe Lenca,
Acioly-Régnier, Pablo Gregori, Elsa Malisani,
Angela Alibrandi, Alain Kuzniak, Rajesh Natajaran,
Jérôme Azé, Eduardo Lacasta, Pilar Orús,
Maurice Bernadet, Dominique Gérard Ramstein,
Julien Blanchard, Lahanier-Reuter, Ansaf Salleb,
Catherine-Marie Chiocca, Stéphane Lallich, Aldo Scimone,
Raphaël Couturier, Letitzia La Tona, Benoît Vaillant,
Stéphane Daviet, Patrick Leconte, Ingrid Verscheure
Jérôme David, Rémi Lehn,

Manuscript coordinator
Bruno Pinaud (LINA, University of Nantes, France)
VIII Preface

Acknowledgments

The editors would like to thank the chapter authors for their insights and
contributions to this book.

The editors would also like to acknowledge the members of the review com-
mittee and the associated referees for their involvement in the review process
of the book, and without whose support the book would not have been satis-
factorily completed.

A special thank goes to H. Briand for his encouragements.

Thanks also to J. Blanchard who has managed the cyberchair web site.

Finally, we thank Springer and the publishing team, and especially T.

Ditzinger and J. Kacprzyk, for their confidence in our project.

Nantes, Régis Gras

December 2007 Einoshin Suzuki
Fabrice Guillet
Filippo Spagnolo
Contents

Introduction
Régis Gras, Einoshin Suzuki, Fabrice Guillet, Filippo Spagnolo . . . . . . . . . 1

Part I Methodology and concepts for SIA

An overview of the Statistical Implicative Analysis (SIA)

development
Régis Gras, Pascale Kuntz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
CHIC:Cohesive Hierarchical Implicative Classification
Raphaël Couturier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Assessing the interestingness of temporal rules with Sequential
Implication Intensity
Julien Blanchard, Fabrice Guillet, Régis Gras . . . . . . . . . . . . . . . . . . . . . . . . 55

Part II Application to concept learning in education, teaching,

and didactics

Student’s Algebraic Knowledge Modelling: Algebraic Context

as Cause of Student’s Actions
Marie-Caroline Croset, Jana Trgalova, Jean-François Nicaud . . . . . . . . . . 75
The graphic illusion of high school students
Eduardo Lacasta, Miguel R. Wilhelmi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Implicative networks of student’s representations of Physical
Activities
Catherine-Marie Chiocca, Ingrid Verscheure . . . . . . . . . . . . . . . . . . . . . . . . . 119
X Contents

A comparison between the hierarchical clustering of variables,

implicative statistical analysis and confirmatory factor
analysis
Iliada Elia, Athanasios Gagatsis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

Implications between learning outcomes in elementary

bayesian inference
Carmen Díaz, Inmaculada de la Fuente, Carmen Batanero . . . . . . . . . . . . 163
Personal Geometrical Working Space: a Didactic
and Statistical Approach
Alain Kuzniak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

Part III A methodological answer in various application

frameworks

Statistical Implicative Analysis of DNA microarrays

Gerard Ramstein . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
On the use of Implication Intensity for matching ontologies
and textual taxonomies
Jérôme David, Fabrice Guillet, Henri Briand, Régis Gras . . . . . . . . . . . . . 227
Modelling by Statistic in Research of Mathematics Education
Elsa Malisani and Aldo Scimone and Filippo Spagnolo . . . . . . . . . . . . . . . . 247
Didactics of Mathematics and Implicative Statistical Analysis
Dominique Lahanier-Reuter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
Using the Statistical Implicative Analysis for Elaborating
Behavioral Referentials
Stéphane Daviet, Fabrice Guillet, Henri Briand, Serge Baquedano,
Vincent Philippé, Régis Gras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299

Fictitious Pupils and Implicative Analysis: a Case Study

Pilar Orús, Pablo Gregori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Identifying didactic and sociocultural obstacles
to conceptualization through Statistical Implicative Analysis
Nadja Maria Acioly-Régnier, Jean-Claude Régnier . . . . . . . . . . . . . . . . . . . . 347

Part IV Extensions to rule interestingness in data mining

Pitfalls for Categorizations of Objective Interestingness

Measures for Rule Discovery
Einoshin Suzuki . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
 Contents XI

Inducing and Evaluating Classification Trees with Statistical

Implicative Criteria
Gilbert Ritschard, Vincent Pisetta, Djamel A. Zighed . . . . . . . . . . . . . . . . . 397
On the behavior of the generalizations of the intensity
of implication: A data-driven comparative study
Benoît Vaillant, Stéphane Lallich, Philippe Lenca . . . . . . . . . . . . . . . . . . . . 421
The TVpercent principle for the counterexamples statistic
Ricco Rakotomalala, Alain Morineau . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
User-System Interaction for Redundancy-Free Knowledge
Discovery in Data
Rémi Lehn, Henri Briand, Fabrice Guillet . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
Fuzzy Knowledge Discovery Based on Statistical Implication
Indexes
Maurice Bernadet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481

About the editors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
List of Contributors

Nadja Maria Acioly-Régnier Raphaël Couturier

EA 3729, University of Lyon, France LIFC, University of Franche-Comte,
[email protected] France
[email protected]
Carmen Batanero
Facultad de Educación, University of
Granada, Spain Marie-Caroline Croset
[email protected] LIG CNRS 5217, University of
Grenoble I, France
Serge Baquedano [email protected]
PerformanSe SAS, Carquefou
(Nantes), France
Carmen Díaz
[email protected]
Facultad de Psicología, University of
Maurice Bernadet Huelva, Spain
LINA CNRS 2729, Polytechnic [email protected]
School of Nantes University, France
[email protected] Jérôme David
LINA CNRS 2729, Polytechnic
Julien Blanchard
School of Nantes University, France
LINA CNRS 2729, Polytechnic
[email protected]
School of Nantes University, France
[email protected]
Stéphane Daviet
Henri Briand LINA CNRS 2729, Polytechnic
LINA CNRS 2729, Polytechnic School of Nantes University, France
School of Nantes University, France [email protected]
[email protected]
Catherine-Marie Chiocca Iliada Elia
ENFA, Castanet-Tolosan (Toulouse), Department of Education, University
France of Cyprus, Cyprus
[email protected] [email protected]
XIV List of Contributors

Inmaculada de la Fuente Stéphane Lallich

Facultad de Psicología, University of ERIC, University of Lyon 2, France
Granada, Spain [email protected]
[email protected]
Rémi Lehn
LINA CNRS 2729, Polytechnic
Athanasios Gagatsis
School of Nantes University, France
Department of Education, University
[email protected]
of Cyprus, Cyprus
[email protected] Philippe Lenca
TAMCIC CNRS 2872, GET/ENST
Régis Gras Bretagne, France
LINA CNRS 2729, Polytechnic [email protected]
School of Nantes University, France
Elsa Malisani
[email protected]
GRIM, University of Palermo, Italy
[email protected]
Fabrice Guillet
LINA CNRS 2729, Polytechnic Alain Morineau
School of Nantes University, France Modulad, Rocquencourt, France
[email protected] [email protected]
Jean-François Nicaud
Pablo Gregori
LIG CNRS 5217, University of
Universitat Jaume I, Castellón,
Grenoble I, France
Spain
[email protected]
[email protected]
Pilar Orús
Pascale Kuntz Universitat Jaume I, Castellón,
LINA CNRS 2729, Polytechnic Spain
School of Nantes University, France [email protected]
[email protected]
Vincent Philippé
PerformanSe SAS, Carquefou
Alain Kuzniak
(Nantes), France
Didirem team, University of Paris 7,
[email protected]
France
[email protected] Vincent Pisetta
ERIC, University of Lyon 2, France
Eduardo Lacasta [email protected]
Departamento de Matemáticas,
Ricco Rakotomalala
Public Universidy of Navarra, Spain
ERIC, University of Lyon 2, France
[email protected]
[email protected]

Dominique Lahanier-Reuter Gérard Ramstein

THEODILE TEAM (E.A. 1764), LINA CNRS 2729, Polytechnic
University of Lille 3, France School of Nantes University, France
[email protected] [email protected]
 List of Contributors XV

Jean-Claude Régnier Jana Trgalova

University of Lyon 2, France LIG CNRS 5217, University of
[email protected] Grenoble I, France
[email protected]
Gilbert Ritschard
Benoît Vaillant
Dept of Econometrics, University of
South Britany University, VALO-
Geneva, Geneva, Switzerland
RIA, France
[email protected]
[email protected]

Aldo Scimone Ingrid Verscheure

GRIM, University of Palermo, Italy LEMME, University of Toulouse 1,
[email protected] France
[email protected]
Filippo Spagnolo
Miguel R. Wilhelmi
GRIM, University of Palermo, Italy
Departamento de Matemáticas,
[email protected]
Universidad Pública de Navarra,
Spain
Einoshin Suzuki [email protected]
Graduate School of Information
Science and Electrical Engineering, Djamel A. Zighed
Kyushu University, Japan ERIC, University of Lyon 2, France
[email protected] [email protected]
Introduction

Régis Gras1 , Einoshin Suzuki2 , Fabrice Guillet1 , and Filippo Spagnolo3

1
LINA, CNRS UMR 6241,
Polytechnic Graduate School of Nantes University, France
[email protected], [email protected]
2
Graduate School of Information Science and Electrical Engineering,
Kyushu University, Japan
[email protected]
3
Department of Mathematics,
University of Palermo, Italy
[email protected]

In the framework of data mining, which has been recognized as one of the ten
emergent technologies for computer sciences, association rule discovery aims
at mining potentially useful implicative patterns from data. Initially stimu-
lated by researches in didactics of mathematics, the Statistical Implicative
Analysis (SIA) offers an original statistical approach based on Implication
Intensity measure, which is dedicated to rule extraction and analysis. Im-
plication Intensity, the first method in SIA in its initial form, evaluates the
interestingness of a rule x → y by the rarity of its number of counter-examples
(xy), according to its probability distribution under an independence hypothe-
sis of x and y. This interestingness measure has been involving a large number
of research works and applications due to its theoretical ethics and practical
merits. Through a graphical interface, CHIC (Cohesive Hierarchical Implica-
tive Classification) software allows easy use of various techniques in SIA for a
wide range of users from experts in data analysis to practitioners with little
background in computer science.
This book includes two complementary topics: on the one hand, theoretical
works related to SIA, or linking SIA with other data analysis methods; and
on the other hand, applied works illustrating the use of SIA in applicative
domains such as: psychology, social sciences, bioinformatics and didactics.
It should be of interest to developers of data mining systems as well as
researchers, students and practitioners devoted to data mining and statistical
data analysis.

R. Gras et al.: Introduction, Studies in Computational Intelligence (SCI) 127, 1–7 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
2 R. Gras, E. Suzuki, F. Guillet and F. Spagnolo

Structure of the book

The book is structured in four parts. The first one gathers three general chap-
ters defining the methodology and the concepts for the Statistical Implicative
Analysis approach. The second part contains six chapters dealing with the use
of SIA as a decision aid tool for the analysis of concept learning in the frame-
work of education, teaching and didactics. In the third part, seven chapters
illustrate the use of SIA as a methodological answer in various application
fields. Lastly, the fourth part includes six chapters describing the extension of
SIA and its application to capture rule interestingness in data mining.

Part I: Methodology and concepts for SIA

• Chapter 1: An overview of the Statistical Implicative Analysis

(SIA) development, by Gras and Kuntz, gives a broad overview of the
Statistical Implicative Analysis which is a data analysis method devoted
to the extraction and the structuration of quasi-implications. It offers a
synthesis which both presents the basic statistical framework of the ap-
proach and details recent developments.

• Chapter 2: CHIC: cohesive hierarchical implicative classification,

by Couturier, is concerned with a data analysis tool based on SIA, named
CHIC. Its aim is to discover relevant implications between variables. It
proposes two different ways to organize these implications into systems: i)
under the form of an oriented hierarchical tree and ii) as an implication
graph. Also, it produces a (non oriented) similarity tree based on the like-
lihood of the links.

• Chapter 3: Assessing the interestingness of temporal rules with

Sequential Implication Intensity, by Blanchard et al., discusses the in-
terestingness of sequential rules which is a key problem in sequence analysis
since the frequent pattern mining algorithms can produce huge amounts
of rules. It defines an original statistical measure named Sequential Impli-
cation Intensity (SII) that evaluates the statistical significance of the rules
according to a probabilistic model. Numerical simulations show that SII
has unique features.

Part II: Application to concept learning in education, teaching

and didactics

• Chapter 4: Student’s Algebraic Knowledge Modelling: Algebraic

Context as Cause of Student’s Actions, by Croset et al.. This chap-
ter describes the construction of a student model in the field of algebra
 Introduction 3

in the framework of Aplusix learning environment. Patterns of student

behaviours are discovered by using SIA. This makes building implicative
connections between algebraic contexts and student’s actions possible.

• Chapter 5: The graphic illusion of high school students, by Lacasta

and Wilhelmi. This chapter deals with the analysis of the relationship be-
tween the mathematical background on linear and quadratic functions,
and the representation of functions (graphics, figures and so on). Factorial
analysis shows a contradiction in the usual assumption of the existence of a
“graphical conceptualization” of functions different from a “non-graphical”
one. Nevertheless, the authors of textbooks and teachers show a trend to
use the graphical representation of functions. In the context of proportion-
ality, the SIA reveals the existence of a graphical illusion shared by high
school students.

• Chapter 6: Implicative networks of student’s representations

of Physical Activities, by Chiocca and Verscheure, proposes to dis-
cuss the results of a questionnaire-based study of young people’s atti-
tudes/representations to team games and volleyball. Processing by CHIC
software shows several networks of variables which make profiling kinds of
students possible. The study of contributions of two additional variables,
sex and gender, enables to improve choices of representative networks for
later interviews. Interestingly and somewhat unexpectedly, while sex is a
strong predictor of attitudes and dispositions to team sports and volley-
ball, gender is not.

• Chapter 7: The structure of conversions among representations

of functions: A comparison between the hierarchical clustering
of variables, implicative statistical analysis and confirmatory fac-
tor analysis, by Elia and Gagatsis, focuses on a comparative study of
three statistical methods, namely the hierarchical clustering of variables,
the implicative method, and the Confirmatory Factor Analysis, applied to
experimental data describing the understanding of functions. The inves-
tigation concentrates on the structure of students’ abilities to carry out
conversions of functions from one mode of representation to another.

• Chapter 8: Implications Between Learning Outcomes in Elemen-

tary Bayesian Inference, by Diaz et al., deals with the use of SIA in
order to study some hypotheses about the interrelationships in students’
understanding of different concepts and procedures after 12 hours of teach-
ing elementary Bayesian inference. A questionnaire made up of 20 multiple
choice items was used to assess learning of 78 psychology students. The
results obtained suggest four groups of interrelated concepts: conditional
probability, logic of statistical inference, probability models and random
4 R. Gras, E. Suzuki, F. Guillet and F. Spagnolo

variables.

• Chapter 9: Personal Geometrical Working Space: a Didactic and

Statistical Approach, by Kuzniak, studies the answers that pre-service
teachers gave in a geometry exercise . The purpose is to improve the un-
derstanding of what we call the geometrical working space. A first study,
based on the notion of geometrical paradigms, leads to a classification of
students’ answers. Then, statistical tools are used in order to fine-tune the
previous analysis and to explain student evolution during their training.

Part III: A methodological answer in various application frame-

works

• Chapter 10: Statistical Implicative Analysis of DNA microar-

rays, by Ramstein, focuses on the application of SIA to microarray gene
expression data. The specificity of these data requires an adaptation of
the concept of intensity of implication. More specifically, it introduces the
concept of rank interval and shows that the integration of the implicative
method in this framework is more efficient than correlation techniques.
The method is applied to the most challenging problems encountered in
gene expression analysis, namely the “discovery of gene coregulation, gene
selection and tumour classification”.

• Chapter 11: On the use of Implication Intensity for matching on-

tologies and textual taxonomies, by David et al., is concerned with the
validation of ontology matching. It is based on an extensional and asym-
metric matching approach designed to find implicative tendencies between
two textual taxonomies or ontologies. More precisely, the chapter focuses
on experimental evaluations of a set of interestingness measures, selected
according to their properties and semantics. The experiments performed
on a benchmark show that the implication intensity delivers the best re-
sults.

• Chapter 12: Modelling by Statistics in Research of Mathematics

Education, by Malisani et al., deals with the theoretical and experimen-
tal relationships between factorial and implicative analyses for modeling
in the framework of didactics of mathematics. A first experiment intro-
duces the supplementary variables for studying some reasoning schemes
on the solution of Goldbach’s conjecture. A second one studies the aspects
of unknown variables and the functional relation in problem-solving in the
contexts of algebra and analytical geometry.

• Chapter 13: Didactics of Mathematics and Implicative Statisti-

cal Analysis, by Lahanier-Reuter, evaluates the assumption that “The
 Introduction 5

Didactics of mathematics has constantly regarded SIA as a profitable and

heuristic method of data analysis”. It shows some explanations such that:
implicative links may be interpreted as rules and regulations connecting
actions, discourses, etc., or as a group’s characteristics. Some examples
show how SIA can be used and what special research results it can pro-
vide. The chapter concludes on some recommendations.

• Chapter 14: Using the Statistical Implicative Analysis for Elab-

orating Behavioral Referentials, by Daviet et al., is concerned with
“PerformanSe Echo” assessment tool that helps human resources managers
in evaluating the behavioral profile of a person along 10 bipolar dimen-
sions. This chapter is interested in building a set of psychological indicators
based on a population of 613 experienced executives who are seeking a job.
The goal is twofold: first to confirm the previous validation study, then to
build a relevant behavioral referential on this population.

• Chapter 15: Fictitious Pupils and Implicative Analysis: a Case

Study, by Orus and Gregori, details a case study, in the context of Di-
dactics of Mathematics, in which they adopt the methodology of using
fictitious data in SIA. Unlike supplementary variables, the fact of adding
fictitious data to the sample does modify the results, so caution is needed.
Nevertheless, fictitious students are a tool for better understanding the
data structure.

• Chapter 16: Identifying didactic and sociocultural obstacles to

conceptualization through Statistical Implicative Analysis, by
Acioly-Reignier and Régnier, aims at understanding the relationship be-
tween culture and cognition. The authors focus on both the roles of written
culture and the teaching and learning strategies involved. The data were
gathered through short interviews and questionnaire based surveys. SIA
enabled them to determine the implicative rules between the responses and
thus the pre-ordering of the responses. The results showed that some spe-
cific symbolic representations constitute didactical and/or socio-cultural
obstacles.

Part IV: Extensions to rule interestingness in data mining

• Chapter 17: Pitfalls for Categorizations of Objective Interesting-

ness Measures for Rule Discovery, by Suzuki, points out four pitfalls
for the categorizations of objective interestingness measures for rule discov-
ery: data bias, rule bias, expert bias, and search bias. The main objective
of this chapter is to issue an alert for the pitfalls which are harmful to one
of the most important research topics in data mining. The author also lists
6 R. Gras, E. Suzuki, F. Guillet and F. Spagnolo

desiderata in categorizing objective interestingness measures.

• Chapter 18: Inducing and Evaluating Classication Trees with

Statistical Implicative Criteria, by Ritschard et al., highlights the in-
terest of SIA for classification trees. It shows how Gras’ implication index
may be defined for rules induced from a decision tree, and that this index
looks like a standardized residual of contingency tables. The first use con-
cerns the a posteriori individual evaluation of the classification rules. The
second use relies on assigning the most appropriate conclusion to each leaf
of the tree. The practical usefulness of this statistical implicative view on
decision trees is demonstrated through a full scale real world application.

• Chapter 19: On the behaviour of the generalisations of the inten-

sity of implication: a data-driven comparative study, by Vaillant
et al., proposes a generalisation of interestingness measure for association
rules, taking into account a reference point, chosen by an expert, in or-
der to apprehend the confidence of a rule. This generalisation introduces
new connections between measures, leads to the enhancement of some of
them, and new parameterised possibilities. The behaviour of the parame-
terised measures is illustrated and discussed on classical datasets. This
study highlights the different properties of each of them and discusses the
advantages of the proposal.

• Chapter 20: The TVpercent principle for the counterexamples

statistic, by Rakotomalala and Morineau, puts into practice the principle
of test value percent criterion for the counterexamples statistic, which is
the basis of SIA approach. It shows how to compute the test value and
what the connection with the measures used by SIA is. The behavior of
these measures is evaluated on a large dataset comprising several hundreds
of thousands of transactions.

• Chapter 21: User-System Interaction for Redundancy-Free

Knowledge Discovery in Data, by Lehn et al., deals with applying
techniques initially designed for redundancy reduction in functional depen-
dencies to association rule reduction. Although the two kinds of relations
have different properties, this method allows very concise representations
that are easily understood by the decider and can be further exploited for
automatic reasoning. This method is compared to other approaches and
tested on synthetic datasets. The information loss resulting from the re-
duction is also discussed.

• Chapter 22: Fuzzy Knowledge Discovery Based on Statistical

Implication Indexes, by Bernadet, is concerned with the application of
SIA to fuzzy knowledge discovery. It explains how to adapt the statisti-
cal indexes to fuzzy knowledge. Yet, these indexes do not evaluate the
 Introduction 7

associated fuzzy rules which depend on the chosen fuzzy operators. The
best fuzzy operators are selected by applying the generalized modus po-
nens on the items of several databases and by comparing its results to
the effective conclusions. By studying methods to aggregate fuzzy rules,
this chapter shows that in order to keep classical reduction schemes,
fuzzy operators must be chosen differently. However, one of these possible
operator sets is also one of the best for processing the generalized modus
ponens.
 Part I

Methodology and concepts for SIA

An overview of the Statistical Implicative
Analysis (SIA) development

Régis Gras and Pascale Kuntz

Laboratoire d’Informatique de Nantes Atlantique

Equipe COnnaissances & Décision
Site Ecole Polytechnique de l’Université de Nantes
La Chantrerie — BP 50609 — 44306 Nantes cedex 3
[email protected], [email protected]

Summary. This paper presents an overview of the Statistical Implicative Analysis

which is a data analysis method devoted to the extraction and the structuration
of quasi-implications. Originally developed by Gras [11] for applications in the di-
dactics of mathematics, it has considerably evolved and has been applied to a wide
range of data, in particular in data mining. This paper is a synthesis which both
briefly presents the basic statistical framework of the approach and details recent
developments.

Key words: quasi-implication, implication intensity, implicative graph, implicative

hierarchy, typicality

1 Introduction
Two important components are involved in the operational human processes of
knowledge acquisition: facts and rules between facts or between rules them-
selves. Through one’s own culture and one’s own personal experience, the
learning process integrates a progressive elaboration of these knowledge forms.
It can be faced with regressions, questions or changes which arise from decisive
quashing, but the knowledge forms contribute to maintain a certain equilib-
rium. The rules formed inductively become quite stable when their success
number -which depends on their explicative or inferential quality- reaches a
certain level of confidence. At first, it is often difficult to replace an initial rule
by another when few counter-examples appear. If they increase, the confidence
in the rule can decrease and the rule can be reajusted or even rejected. How-
ever, when confirmations are numerous and counter-examples are rare, the
rule is robust and can stay in our minds. For instance, let us consider the
acceptable rule “All Ferraris are red”. Even if one or two counter-examples
happen this rule is maintained, and it will be even confirmed again by new
examples.
R. Gras and P. Kuntz: An overview of the Statistical Implicative Analysis (SIA) development,
Studies in Computational Intelligence (SCI) 127, 11–40 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
12 R. Gras, P. Kuntz

Hence, contrary to what happens in mathematics where rules do not allow

for any exception, the rules considered in human sciences are considered to
be acceptable when the number of counter-examples remains “tolerable” in
view of the number of situations where they are positive and efficient. In data
analysis, the problem is to determine a consensus criterion which quantifies
the confidence quality level of the rule according to the user’s requirements.
Our approach rests on three epistemological assumptions. The criterion is
statistical. It is non linearly, robust to noise (i.e. not very influenced by the
first counter-examples), and it becomes very low if the counter-examples often
reappear. Our choice can be questioned, however it has been confirmed in
various situations.

1.1 From didactics to data mining

“If a question is more complex than another, then each pupil who succeeds
in the first one should also succeed in the second one”. Every teacher knows
that this situation shows exceptions whatever the complexity degree between
questions. The evaluation and the structuration of such implicative relation-
ships between didactic situations are the generic problems at the origin of the
development of the Statistical Implicative Analysis (SIA) [11]. These prob-
lems, which have also drawn attention from psychologists interested in ability
tests [5, 27], have known a significant renewed interest in the last decade in
data mining.
Indeed, quasi-implications, also called association rules in this field, have
become the major concept in data mining to represent implicative trends be-
tween itemset patterns. In data mining, the paradigmatic framework is the
so-called basket analysis where a quasi-implication Ti → Tj means that if a
transaction contains a set of items Ti then it is likely to contain a set of items
Tj too. For simplicity’s sake, let us now on call “rule” a quasi-implication.
In data mining, rules are computed on large size databases. From the sem-
inal work of Agrawal et al. [1], numerous algorithms have been proposed to
mine such rules. Most of them attempt to extract a restricted set of relevant
rules, easy to interpret for decision-making. Yet, comparative experiments
have shown that results may vary with the choice of rule quality measures
(e.g. [13, 25]). In the rich literature devoted to this problem, interestingness
measures are often classified into two categories: the subjective (user-driven)
ones and the objective (data-driven) ones. Subjective measures aim at taking
into account unexpectedness and actionability relatively to prior knowledge,
while objective measures give priority to statistical criteria. Among the latter,
the most commonly used criterion for quantifiying the quality of a rule a → b
is the combination of the support (the frequency f (a ∧ b)) which indicates
whether the items a and b occur reasonably often in the database, with the
confidence (the conditional frequency). However, it is well-known that the
confidence presents a major default: it is insensitive to the dilatation of f (a),
f (b) and the database size. Other functions measure a link or an absence of
 An overview of the Statistical Implicative Analysis (SIA) development 13

link between the items but, like χ2 , they do not clearly specify the direction
of the relationship. Moreover, in addition to rule filtering, rule structuring is
necessary to highlight relationships and makes rule interpretation both easier
and more accurate.
The SIA provides a complete framework to evaluate the interestingness of
the rules and to structure them in order to discover relationships at differ-
ent granularity levels. The underlying objective is to highlight the emerging
properties of the whole system which can not be deduced from a simple de-
composition into sub-parts (e.g. [30]). All these properties, which emerge from
complex interactions -probably non linear-, contribute to the interpretation
of the global nature of the system.

1.2 Contents of the paper

Section 2 presents the statistical framework to measure the rule quality: we

first remind the reader the definition of the implication intensity for binary
variables and propose different properties. Section 3 presents the extensions of
the basic definition for different types of variables (modal, frequential, inter-
val), and an entropic version adapted to large datasets. The following sections
are concerned with rule structuration. Section 4 defines the implicative graph.
Section 5 generalizes the notion of rule to the notion of R-rule (rule of rule),
and section 6 describes the combinatorial structure of an implicative hierarchy
whose elements are R-rules. Aids for analyzing these complex structures are
developed in section 7 (significative levels of the implicative hierarchy) and
section 8 (supplementary individuals and variables). An illustration from a
real data corpus coming from a survey on teacher’s perception of training in
mathematics is presented in section 9.

2 The implicative intensity for the binary case

2.1 The basic situation

Let us consider a population E of n objects or individuals described by a

finite set V of binary variables (attributes, criteria, scores, . . . ). We are here
interested in the following question: “To what extent the variable b is true when
the variable a is true” ? In other words, “do the subjects have a tendency for
having b when we know that they have a? ” In real-life situations — e.g. in
human sciences— deductive theorems of the logical form a ⇒ b are often
difficult to establish because of the exceptions. Consequently, it is necessary
to “mine” the dataset to extract rules reliable enough to conjecture causal
relationships which structure the population. At the descriptive level, they
allow to detect a certain stability in the structuration. And, at the predictive
level, they allow to make assumptions. However, the rule mining processes
require rigorous approaches which prevent a too flimsy empiricism.
14 R. Gras, P. Kuntz

2.2 The statistical framework

Our approach, based on the non-parametric test reasoning, is close to the

Likelihood Linkage Analysis (LLA) developed by I.C. Lerman [26]. The quality
measurement of an implicative relationship a → b is based on the unlikelihood
of the counter-example number in the dataset i.e. cases where b is false when
a is true [11, 12, 14]. To quantify this unlikelihood, we compare the deviation
between the contingency and a theoretical model associated with a random
drawing. In exploratory data analysis, we consider the deviation value and not
just the H0 acceptation/reject. This measure quantifies the “surprisingness”
of the expert faced with a number of counter-examples improbably small for
an independence assumed between the variables and for the cardinalities of
the considered data.
More precisely, let us denote by A ⊂ E the subset of individuals for which
a is true, by A its complementary set and by na = card (A) (resp. na ) the
cardinal of A (resp. A). The logical rule A ⇒ B is true when A ⊂ B. However,
this strict inclusion is exceptionally observed in real-life situations; in practice,
it is quite common to observe a few subjects where a is true and b is false
without having the general trend to have b when a is true contested. Conse-
quently, we consider in the following quasi-rules —called rules for simplicity’s
sake— of the form a → b.

2.3 Definitions

To accept or
reject a → b it is quite common to consider the number na∧b =

card A ∩ B of counter-examples. However, to quantify the surprisingness of
the rule, this number must be relativized according to n, na and nb . Intuitively,
it is all the more surprising to discover that a rule has a small number of
counter-examples as the data set is large. The objective of the implicative
intensity is precisely to express the unlikelihood of na∧b in E.
We compare the observed number of counter-examples na∧b with the num-
ber of expected counter-examples for an independence hypothesis. Like I.C.
Lerman with the similarity in LLA [26], we randomly draw two subsets X and
Y of, respectively, na and nb elements.

Definition 1. The rule a → b is said to be admissible for a given thresh-

old α if the
probability of having the observed number of counter-examples


card A ∩ B greater than the expected number card X ∩ Y is smaller
than α:



Pr card X ∩ Y ≤ card A ∩ B ≤ α


The distribution of card X ∩ Y depends on the drawing pattern. When
X and Y are draw with throw-in the distribution is Binomial, otherwise it is
Hypergeometric.
 An overview of the Statistical Implicative Analysis (SIA) development 15


Remark 1. For a certain process of drawing, the random variable card X ∩ Y
n n
follows a Poissonian distribution P (λ) with λ = an b .
Let us consider a process where the individuals dynamically arrive e.g. a
flow of transactions which fill up a database. We stop the process when there

are na individuals with a true and nb individuals with b true. Let card X ∩ Y
be the random variable associated with the counter-example number during
the process. We suppose that the process checks three hypotheses : (i) the
waiting times for the events (a and b) are independent random variables, (ii)
the distribution of the number of events which happen in the interval [t, t + T ]
only depends on T , (iii) two events can not simultaneously happen.
Consequently, the number of events which happen during a fixed period
follows a Poissonian distribution P (λ) where λ is the cadence of the event
arrival.
The probability of the event (a = 1) (resp. (b = 0)) is estimated by the
frequency na /n (resp. nb /n). Then, the probability of the joint event (a = 1
and b = 0) is estimated by
na nb
.
n n
Hence, for a flow of n individual, the arrivals of the event (a = 1 and b = 0)
n n
follow a Poissonian distribution with parameter λ = an b .
Consequently,
λs
Pr card X ∩ Y = s = e−λ



s!
and the probability that the chance leads to a greater number of counter-
examples than those observed is defined by
card(A∩B )



X λs −λ
Pr card X ∩ Y ≤ card A ∩ B = e
s=0
s!

In the following, we consider the Poissonian distribution. In the classical

approximation conditions, the other distributions converge on the Poissonian
type.

Let us consider, for nb 6= 0, the standardized random variable Q a, b :

card X ∩ Y − nannb


Q a, b = q
na nb
n



We denote by q a, b the observed value of Q a, b in the experimental
realization. It is defined by
na nb

n −
q a, b = a∧b
q n
na nb
n
16 R. Gras, P. Kuntz

This value measures a deviation between the contingency and the expected
value when a and b are independent.

the approximation is justified (e.g. λ > 4) the random variable
When
Q a, b is approximatively N (0, 1)-distributed.

Definition 2. The implication intensity ϕ (a, b) of the rule a → b is defined by

Z ∞
1 t2
e− 2 dt




ϕ (a, b) = 1 − Pr Q a, b ≤ q a, b = √
2π q(a,b)

if nb 6= n, and ϕ (a, b) = 0 otherwise.

Definition 3. The implication intensity ϕ (a, b) is admissible for a given

threshold α if ϕ (a, b) ≥ 1 − α.
The implication intensity measures the surprisingness to observe a small
number of counter-examples. It is an inductive and informative quality mea-
sure. Consequently, if the rule is trivial —-i.e. when B is small or equal to
E— this surprisingness is small.

Proposition 1. [12] Let us suppose that na is fixed and A ⊂ B. If nb tends

towards n, then ϕ (a, b) tends towards 0.
We set ϕ (a, b) = 0 if nb = n by continuity (consequence of the property 1).
If A ⊂ B then ϕ (a, b) can be smaller than 1 when the surprisingness is not
sufficient.

2.4 Comparison with some classical measures

The observed quasi-implication q a, b is not symmetrical. It is different from
the Pearson’s correlation coefficient ρ (a, b) which measures the linkage be-
tween a and b.

Proposition 2. [12] Let ρ (a, b) be the value
of the Pearson’s correlation

between the binary variables a and b. If q a, b 6= 0 then
r
ρ (a, b) n

=−
q a, b nb na
The variation of the implication intensity is different from the Loevinger’s
coefficient [27] and from the confidence conf (a, b) = na∧b /na . It increases non
linearly with the increasing of E, A and B, and it decreases with the trivial
situations. Moreover, the maximal intensity is not necessarily reached for the
inclusion A ⊂ B; indeed, the inductive quality may be quite low, whereas
conf (a, b) = 1 [13].
 An overview of the Statistical Implicative Analysis (SIA) development 17

2.5 Stability of the implication intensity

The implication intensity is noise-resistant in the neighbourhood of na∧b =

0 [13].
In the following, we study the sensitivity of ϕ (a, b) for small variations
of the parameters n, na , nb and na∧b . Previous numerical experiments have
confirmed the influence of the parameter variations on ϕ [10, 13]. Here, we
study the differentiation of q.
Let us consider the parameters n, na , nb and na∧b as real numbers which
satisfy the following inequalities: na∧b ≤ inf (na , nb ) et sup (na , nb ) ≤ n. In
this case, q can be considered as a continuous differentiable function:
∂q ∂q ∂q ∂q
dq = dn + dna + dnb + dn
∂n ∂na ∂nb ∂na∧b a∧b

To study the variability of q depending on nb , we replace nb by n − nb ,

and consequently the sign in the partial derivative.

Example 2. Let us suppose that na is constant, and that nb and na∧b

may vary. Then,
∂q na 1/2 −3/2 na 1/2 −1/2
= 12 na∧b 1



∂nb n (n − nb ) + 2 n (n − nb )
∂q q 1
∂n = n
a∧b a∧b
n
∂q
∂na =0


Consequently, if ∆nb and ∆na∧b are positive, then ∆q a, b is positive.
This property can be interpreted as follows: for fixed n and na , the implica-
tion intensity decreases when the numbers of the b examples and the a ⇒ b
counter-examples increase. The implication intensity is maximal for the ob-
served values nb and na∧b , and minimal for nb + ∆nb and na∧b + ∆na∧b .
To examine the sensibility of the implication intensity, we consider ϕ as a
function of q: Z ∞
1 2
ϕ (q) = √ e−t dt
2π q
By differentiation, we obtain
dϕ 1 2
= − √ e−q < 0
dq 2π
This result confirms that the implication intensity decreases with q, and
it gives the speed of the variation.
With a similar approach, let us compare the stability of ϕ with the stability
of the confidence conf (a, b). The sensibility of conf to the variation of the
counter-examples is defined by
18 R. Gras, P. Kuntz

∂c 1
=−
∂na∧b na
Consequently, as expected, the confidence increases when na∧b decreases.
However, the variation of the decreasing speed is constant whatever n and nb .
This situation highlights the limits of the parameter role in the sensitivity of
the measure.

3 Extensions to different types of variables

3.1 Modal and frequential variables
The basic situation
The first applicative framework of this research was concerned with the rep-
resentation that the teachers have of their own practice [3]. In a survey, a set
of teachers has been asked to order a list of significative words depending on
their importance. The resulting implications were: “if I select a word x with
the importance ix then I select the word y with the importance iy ≥ ix ”.
In this case, we consider modal variables a ∈ [0, 1] which describe satisfac-
tion degrees. A similar case appears in situations where the variable frequency
can be interpreted as a pre-order on the set of the values given by the subjects.
Such situation appears in didactics when we study the success frequency for
a test composed of questions coming from different domains.

Formalization
Let us denote by a (i) and b (i) the values of i ∈ E for the modal variables a
and b, and by sa and sb their empirical standard-deviations.

Definition 4. [24]. For a pair (a, b) of modal variables, the implication in-
tensity, called the propension index, is defined by
P n n
a (i) b (i) − an b
qp a, b = qi∈E


(n2 s2a +n2a )(n2 s2 +n2 )
b b
n3



Proposition 3. When a and b are binary variables then qp a, b = q a, b .
In this case, it is easy to prove that n2 s2a + n2a = nna , n2 sb2 + n2b = nnb
P
and i∈E a (i) b (i) = na∧b .
This extension remains valid for the frequential variables and the positive
numerical variables when they are normalized: e a (i) = a (i) / maxi∈E a (i).
A similar measure has been recently introduced by Régnier and Gras [29]
for ranking variables associated with a total order on a set of choices presented
to a judge population. In this case, the considered implication is “if an object
i is ordered by the judges at a place pi then an object j is ordered by the
same judges at a place pj > pi ”.
 An overview of the Statistical Implicative Analysis (SIA) development 19

3.2 Variables on intervals

The basic situation

Let us consider a given set of biometric data. The considered implication is

“if the weight of a male is between 65 and 70 kgs then his height is between
1.70 and 1.76 m”.
More generally, let us consider two real variables a and b with a finite
number of values in the respective intervals A = [a1 , a2 ] and B = [b1 , b2 ].
Roughly speaking, the problem consists in finding implicative trends between
representative unions of sub-intervals of A and B.

Main steps of the heuristic

The problem is decomposed in two steps. First, we partition the inter-

vals A and B in a finite number of sub-intervals {A1 , A2 , . . . , Ap } and
{B1 , B2 , . . . , Bq } which depend on the structure of the a and b distributions:
there is an internal statistical homogeneity in each Ai (resp. Bi ) and a high
dispersion between each pair Ai , Aj (resp. Bi , Bj ). Second, we compute the
most significative implicative trends between unions of Ai and unions of Bj .
We have adapted the k-means algorithm for the interval partitioning prob-
lem [16]. The quality criteria of the partition are the intra-class and the inter-
class inertia. Let π (A) and π (B) be two partitions obtained by this approach
which respectively contain nA and nB elements. We denote by Ω (π (A)) (resp.
Ω (π (B))) the set of the 2nA−1 (resp. 2nB −1 ) partitions of A (resp. B) com-
posed of the unions of elements of π (A) (resp. π (B)) associated with adja-
cent intervals in A (resp. in B). For instance, if π (A) = {A1 , A2 , A3 , A4 } s.t.
A = A1 ∪ A2 ∪ A3 ∪ A4 then

{{{A1 } , {A2 } , {A3 } , {A4 }} , {{A1 A2 } , {A3 } , {A4 }} , . . . ,

Ω (π (A)) =
{A1 A2 A3 A4 }}

For each pair (Pi , Pj ) ∈ Ω (π (A)) × Ω (π (B)) (resp. (Pj , Pi ) ∈ Ω (π (B)) ×

Ω (π (A))) we compute the geometric mean of the implication intensities be-
tween each sub-interval of Pi (resp. Pj ) and each sub-interval of Pj (resp. Pj ).
Let us denote by maxAB and maxBA the respective maximal values between
Ω (π (A)) and Ω (π (B)) and between Ω (π (B)) and Ω (π (A)). The implica-
tion is optimal if there is a partitioning of A which corresponds to maxAB
and a partitioning of interval of B which corresponds to maxBA .

3.3 Interval variables

The basic situation

Let us consider a score distribution of a class for different subjects. The consid-
ered implication is “the sub-interval [2; 5.5] in mathematics generally implies
20 R. Gras, P. Kuntz

the sub-interval [4.25; 7.5] in physics”. These two sub-intervals belong to an

“optimal” partition -according to the inertia- of the definition domains [1; 18]
and [3; 20] of the scores in mathematics and in physics.

Main steps of the heuristic

The previous approach can be adapted to the interval variables, which are
symbolic data. Let us consider two variables a and b which are associated
with a series of intervals due to the measure imprecision: Iia (resp. Iib ) is the
interval of a (resp. b) for the individual i ∈ E. Let I a (resp. I b ) be the interval
which contains all the a (resp. b) values. We can define on I a and I b a partition
which optimizes a given criterium. The intersections between Iia and I a and
between Iib and I b follow a distribution that takes into account the common
parts. Consequently, the problem is similar to the computation of the rules
between the on-interval variables (we refer to [16] for details).

3.4 The entropic version of the implication intensity

The limits of the basic implication intensity for large datasets

Pertinent results have been obtained with the implicative intensity ϕ for var-
ious applications where the data corpuses are relatively small (n < 300).
However, in data mining, numerical experiments have highlighted two limits
of ϕ for large datasets. First, it tends to be not discriminant enough when
the size of E dramatically increases (e.g. [8]); its values are close to 1 even
though the inclusion A ⊂ B is far from being perfect. Second, like numerous
measures proposed in the literature, it does not take into account the con-
trapositive b ⇒ a which could allow to reinforce the affirmation of the good
quality of the implicative relationship between a and b, and the capacity to
estimate the causality between the variables.

The entropic implication intensity

To overcome these difficulties, we have proposed to modulate the value of

the surprise quantified by the implication intensity by taking
into account
both the imbalance between card (A ∩ B) and card A ∩ B associated with



a ⇒ b and the imbalance between card A ∩ B and card A ∩ B associated
with the contrapositive b ⇒ a [6, 7, 18]. We have introduced a new measure,
called the entropic implication intensity, based on the Shannon’s entropy to
non-linearly quantify these differences.
More precisely, let us first consider a weighted version of the implication
1/2
intensity φ (a, b) = (ϕ (a, b) .τ (a, b)) where τ (a, b) measures the imbalance
between na∧b and na∧b and the imbalance between na∧b and na∧b . Intuitively,
the surprise measured by φ must be softened (resp. confirmed) when the
 An overview of the Statistical Implicative Analysis (SIA) development 21

number of counter-examples na∧b is high (resp. small) for the rule and its
contrapositive considering the observed numbers na and nb .
A well-known index for taking the imbalances into account non-linearly is
the Shannon’s conditional entropy. The conditional entropy Hb/a of cases (a
and b) and (a and b) given a is defined by
na∧b na∧b n n
Hb/a = − log2 − a∧b log2 a∧b
na na na na

and, similarly, the conditional entropy Hb/a of cases (a and b) and (a and
b) given b is defined by
na∧b n n n
Ha/b = − log2 a∧b − a∧b log2 a∧b
nb nb nb nb

We can here consider that these entropies measure the average uncertainty
on the random experiments in which we check whether b (resp. a) is realized
when a (resp. b) is observed. The complements of 1 for these uncertainties
Ib/a = 1 − Hb/a and Ib/a = 1 − Hb/a can be interpreted as the average
information collected by the realization of these experiments; the higher this
information is, the stronger the guarantee of the quality of the implication
and its contrapositive will be.
Intuitively, the expected behavior of the measure φ is determined by three
phases:
1. a slow reaction to the first counter-examples (robustness to noise).
2. an acceleration of the reject in the neighborhood of the balance.
3. an increasing rejection beyond the balance -which was not guaranteed by
the basic implication intensity ϕ.
Hence, in order to have the expected significance, our model must satisfy the
following constraints:
1. Integrating both the information relative to a → b and that relative to
b → a respectively measured by Ib/a and Ia/b . A product Ib/a .Ia/b is
well-adapted to simultaneously highlight the quality of these two values.
2. Raising the conditional entropies to the power for a fixed number α > 1
in the information definitionsto reinforce
 the
 contrast
between the differ-
1/β
α α
ent phases described below: 1 − Hb/a . 1 − Ha/b with β = 2α to
remain of the same dimension as ϕ.
3. The need to consider that the implications have lost their inclusive mean-
ing when the number of counter-examples is greater than half of the ob-
servation
 of a and
  b. Beyond
 these values we consider that the terms
α α
1 − Hb/a and 1 − Hb/a are equal to 0.
22 R. Gras, P. Kuntz
n
Let fa = nna (resp. fb = nb ) be the frequency of a (resp. b) on E and fa∧b
be the frequency of the counter-examples. The proposed adjustment of the
previous informations Ib/a and Ia/b can be defined by-
     α
α f f f fa∧b
α
Ic b/a = 1−Hb/a = 1+ 1 − a∧b log2 1 − a∧b + a∧b log2
fa fa fa fa
h h
if fa∧b ∈ 0, f2a ; otherwise, Ic
α
a/b = 0
and
     α
α α fa∧b fa∧b fa∧b fa∧b
I a/b = 1−Hb/a = 1+
c 1− log2 1 − + log2
fb fb fb fb
h h
if fa∧b ∈ 0, f2a ; otherwise, Ic
α
a/b = 0.

Definition 5. The imbalances are measured by τ (a, b) —called the inclusion

index— defined by
 1/2α
α
τ (a, b) = Ic α
b/a .I a/b
c

and, the weighted version of the implication intensity —called the entropic
implication intensity— is given by
1/2
φ (a, b) = (ϕ (a, b) .τ (a, b))

Example 3.

P
b b
a 200 400 600
a 600 2800 3400
P
800 3200 4000
P
b b
a 400 200 600
a 1000 2400 3400
P
1400 2600 4000
P
b b
a 40 20 60
a 60 280 340
P
100 300 400

Table 1. Distribution examples (a, b and c).

For the table 1.a, the implicative intensity is ϕ (a, b) = 0.9999. The entropic
functions are Ha/b = 0 = Hb/a . The weighting coefficient is τ (a, b) = 0. And,
 An overview of the Statistical Implicative Analysis (SIA) development 23

φ (a, b) = 0 whereas the confidence c (a, b) is equal to 0.333. The entropic

functions moderate the implication intensity when the inclusion is bad.
For the table 1.b, the implication intensity ϕ (a, b) = 1. The entropic
functions are Ha/b = 0.918 and Hb/a = 0.391. The weighting coefficient is
τ (a, b) = 0.6035. And, φ (a, b) = 0.777 and the confidence c (a, b) = 0.666.
The table 1.c proves that the correspondance between ϕ and φ is not
monotonous. The intensity implication is lower for the table 1.c than for the
table 1.b. And, it is the contrary for φ (a, b). Let us remark that the confidence
is the same for the two tables.

4 The implicative graph

When computing the implication intensities between all pairs of variables of V ,
we obtain a square matrix M of numbers in [0, 1]. The global structure of the
relationships between the variables does not clearly appear. To highlight this
structure we have associated a directed graph with M , called the implicative
graph [2, 11].
Let Φα be the relationship defined on V × V by ϕ for a given threshold
α ∈ [0, 1]: aΦα b if and only if ϕ (a, b)≥ α. The threshold α, which controls the
implicative quality of the rules, is chosen by the user. The relationship Φα is
reflexive, not symmetric and not transitive. However, it is interesting to con-
sider the partial order relationships between the subsets of V . Consequently,
we extend the relationship Φα : if aΦα b and bΦα c then we accept the transitive
closure aΦα c if and only if ϕ (a, c) ≥ 0.5 i.e. when the implicative trend of a
on c is better than the neutrality.
Hence, for a given threshold α, the graph GM,α is defined as follows: its
vertices are the variables of V , and there is an arc between a pair of variables
(a, b) if and only if aΦα b.
Different options of the software CHIC allows to easily interact with the
drawing of the graph.

5 From rules to R-rules

5.1 The basic situation

In the didactics of mathematics, one of the fundamental question is to identify

the source of the problems -both didactical and epistemological- the pupil is
faced with during his learning processes. These obstacles are based on the con-
ceptions the pupil is building up. These conceptions are structured by simple
or complex rules which together allow to elaborate the basis of a cognitive
model.
This structuration is neither a simple union of rules nor a classical
hierarchical structure where the variable classes are fit into partitions which
24 R. Gras, P. Kuntz

are partially ordered by the relation “thiner than” which reflects the simi-
larity between the class elements. To complete the information provided by
the previous models, we have proposed the concept of R-rules (rules of rules)
which are an extension of the quasi-implications: their premisses and their
conclusions can be rules themselves [15, 17, 20, 23]. To guide the intuition a
parallel can be drawn from the proof theory with the logical implication:
(X ⇒ Y ) ⇒ (Z ⇒ W ) describes an implication between the two theorems
X ⇒ Y and Z ⇒ W previously established.

5.2 The R-rules and their interpretation

In the following, we consider binary variables. The R-rules are an extension

of the classical binary rules a → b to rules of rules R0 → R00 , which may be
complex themselves. For instance a → (b → c) is a R-rule between a variable
a and a rule (b → c), and (a → b) → (c → d) is a R-rule between two rules
(a → b) and (c → d). To indicate the complexity of the implication composi-
tion, we associate a complexity degree with each R-rule.

Definition 6. The R-rules of degree 0 are variables of V . The R-rules of

degree 1 are the simple quasi-implications of the form a → b. A R-rule of
degree i, 1 < i ≤ p, is a rule R0 → R00 between two R-rules R0 and R00 whose
respective degrees j and k satisfy j + k = i − 1.

For instance, a → b is a R-rule of degree 1, a → (b → c) a R-rule of degree

2 and (a → b) → (c → d) a R-rule of degree 3. When there is no ambiguity
we denote by R a R-rule of degree greater or equal than 1.
The R-rules allow to express different levels of abstraction: (1) situation
or object descriptions (conjunction of R-rules of degree 0), (2) implications
between variables (R-rules of degree 1), and (3) implications between impli-
cations (some R-rules of degree greater than 1). Consequently, their interpre-
tation may vary according to three typical cases:

1. when R → a then a may be interpreted as a quasi-consequence of R;

2. the R-rule a → R means that a R-rule R may be partially deduced
from the observation of a. Moreover, although we here consider quasi-
implications only, the intuition can be supported by Heyting algebra where
an implication a ⇒ (b ⇒ c) is equivalent to (a AN D b) ⇒ c;
3. the R-rule R0 → R00 means that the property R00 is the quasi-corollary of
a previous property R0

5.3 A measure of cohesion of the R-rules

The objective is to discover R-rules with a good implicative quality —called

cohesion in the following— i.e. R-rules R0 → R00 with a strong implicative
 An overview of the Statistical Implicative Analysis (SIA) development 25

relationship between the components of R0 and those of R00 . For instance, it

seems natural to form a R-rule (a → b) → (c → d) if the implicative relation-
ships a → c, a → d, b → c and b → d are significant enough. Intuitively, this
means that they must contrast with the disorder of a random experience.
The entropy is well-suited to measure this disorder. Let us first consider
a R-rule a → b of degree
1. And,

let Y be the random indicator vari-
able of the event Q a, b ≥ q a, b . The distribution of Y is defined by
Pr (Y = 1) = ϕ (a, b) and Pr (Y = 0) = 1 − ϕ (a, b). The entropy of this expe-
rience is −p log2 p − (1 − p) log2 (1 − p) where p = ϕ (a, b).
The extreme values are 0 if ϕ (a, b) = 0 (by setting 0 log2 0 = 0) and 1 if
ϕ (a, b) = 0.5. This last value is reached when na∧b = na nb /n i.e. when na∧b
is equal to the expected mean. In this case, when ϕ (a, b) < 0.5, the mean-
ing of the implication is lost and it seems natural to set te cohesion equal to 0.

Definition 7. The cohesion c (a, b) of a R-rule a → b of degree 1 is defined by

 1/2
2
c (a, b) = 1 − (−p log2 p − (1 − p) log2 (1 − p))

if p = ϕ (a, b) > 0.5 and c (a, b) = 0 otherwise.

We square the entropy to reinforce the contrast between values in [0, 1]
and the square root to the complement to 1 allows to measure the cohesion
on a same scale as the entropy.
The generalization of this definition to R-rules of higher degree is guided
by the following requirement: the cohesion of R0 → R00 must take into
account both the cohesion of R0 and R00 as well as the implicative relation-
ships between the attributes of R0 and those of R00 . Let ≺R be the left right
reading order on the variables which composed a R-rule. For instance, for
(a → b) → (c → d) the order on {a, b, c, d} is defined by a ≺R b ≺R c ≺R d.
Then, a simple way to satisfy the previous requirements is to take the
mean of the cohesions of R0 and R00 and of the cohesions if each ordered pairs
composed of one attribute of R0 and one attribute of R00 in accordance with
the permutation orders. Here we favour the geometric mean as it is equal to
0 as soon as the cohesion of one ordered pair is equal to 0 (i.e. when an im-
plication is low or without surprise) and it is close to 1 when the cohesions of
all ordered pairs are high.

Definition 8. Let R be a R-rule of the form R0 → R00 where R0 and R00

are respectively associated with the orders a01 ≺R0 a02 ≺R0 . . . a0k and a001 ≺R00
a002 ≺R00 . . . a00h . The cohesion of R is defined by
 2/r(r−1)
     
0 0 00 0 00
Y Y Y
c (R) =  c ai , aj . c ai , aj . c ai , aj 
i=1,k−1;j=2,k i=1,h−1;j=2,h i=1,k;j=1,h

where r = k + h.
26 R. Gras, P. Kuntz

6 The implicative hierarchy

6.1 The basic situation

Generally speaking, R-rules contribute to increasing the analysis richness. We

do not solely extract facts or isolated behaviors, but more general conducts,
revealing more global, less singular phenomena i.e. in didactics profound psy-
chological representations. The different complexity degrees of the R-rules
can be associated with a hierarchical structure which reflects the genesis of
the “operating knowledge” developed by Piaget [28]. We go from one level to
another by a process of reflecting abstraction: from object representation to
representation of operations on the objects, then to representation of opera-
tions on the operations. This process involves a dynamical hierarchical point
of view in contrast with the static point of view associated with a taxonomy.
Hence, the individual description of the R-rules by aggregating simple rules
is not sufficient. It is necessary to develop a global structure which reflects
the emerging properties of the whole. Consequently, we have developed the
concept of “implicative hierarchy” to structure the significant R-rules.
Let us introduce this notion by an example. A graphical representation
of an implicative hierarchy on the variable set V = {a, b, c, d, e} is given on
→
−
figure 1. The elements of the implicative hierarchy H V are R-rules:
−→
HV = {a, b, c, d, e, b → c, e → d, a → (e → d)}

a e d b c
→
−
Fig. 1. Graphical representation of the implicative hierarchy H V = {a, b, c, d,
e, b → c, e → d, a → (e → d)}

Note that contrary to hierarchies in classical hierarchical classification

(HC) the tree associated with the implicative hierarchy is not necessarily
connected. Intuitively, this means that it contains only significant R-rules ac-
cording to the cohesion measure.

6.2 Definitions

The R-rules which composed an implicative hierarchy can be associated with

k-permutations —called classes by analogy with the HC— that satisfy special
 An overview of the Statistical Implicative Analysis (SIA) development 27

interlocking conditions. For instance, in the example given below, the R-rule
a → (e → d) is associated with the permutation aed. And, this is the only
possible association as the R-rule (a → e) associated with the permutation ae
→
− →
−
is not in H V . The class set HV associated with H V is

HV = {a, b, c, d, e, bc, ed, aed}

The R-rules are deduced by a recursive decomposition of the non elemen-

tary classes of HV . The class aed is the unique amalgamation of a ∈ HV
and ed ∈ HV . Since the class ed is associated with e → d, the class aed is
associated with a → (e → d).
More formally, let ΩV be the set of all k-permutations on the variable set
V , k = 1, p. The elements C of ΩV are strings with distinct characters. Let
≺ be the left-right reading order on the variables of a permutation of ΩV as
we defined it previously. In order to compare and combine the elements of
ΩV to form an implicative hierarchy, we define three operators on ΩV , whose
appelations are inspired by the set theory:
• Intersection. The intersection C 0 ∩ b C 00 of two strings of ΩV is the largest
sub-string of contigous variables common to C 0 and C 00 . In case of equality
we keep the first sub-string of C 0 according to ≺. If C 0 = acdb and C 00 =
cdab then C 0 ∩b C 00 = cd, and if C 0 = abcd and C 00 = cdab then C 0 ∩ b C 00 = ab.
• Union. The union C ∪C of two distinct strings C and C s.t. C 0 ∩
0 b 00 0 00 b C 00 = ∅
is the concatenation of C and C with C first according to ≺. If C 0 = aceb
0 00 0

and C 00 = f gh than C 0 ∪ b C 00 = acebf gh.

• Difference. For three strings C, C 0 and C 00 of ΩV s.t. C = C 0 ∪ b C 00 , the dif-
0 0 00 b 00 between
ference C −C
b between C and C is C and the difference C −C
00 0 0 00 b 0 = c and
C and C is C . If C = abc, C = ab and C = c then C −C
00
C −C
b = ab.

Definition 9. An implicative hierarchy HV is a subset of permutations of ΩV

satisfying the three following requirements:
1. HV contains the variables of V , called elementary classes
2. for each pair C 0 , C 00 ∈ HV , C 0 ∩
b C 00 = {∅, C 0 , C 00 }
3. for each non elementary class C ∈ HV , there is a single pair C 0 , C 00 ∈ HV
s.t. C = C 0 ∪
b C 00

From the condition 2, a hierarchy is a partially ordered set with the in-
b defined on ΩV by: C 0 ⊂C
clusion relation ⊂ b 00 if and only if C 0 ∩
b C 00 = C 0 . The
condition 3 is required to recover all the classes of the hierarchy.
The isolated interpretation of a class of the hierarchy is tricky since it is
a k-permutation which does not state the implication composition. For in-
stance, if we analyse the class aed ∈ HV all alone, we do not know the exact
meaning of aed: it could be either a → (e → d) or (a → e) → d. However, the
28 R. Gras, P. Kuntz

whole HV class set allows to dispel ambiguity: a → (e → d) is chosen as ed is

a class of HV .

Proposition 4. [15] Each non elementary class C of an implicative hierarchy

HV can be associated with a unique R-rule.
→
−
The R-rule set H V associated with HV can be graphically represented by
a valuated binary directed tree:

• each of the elementary classes are located at a terminal node;

• each of the internal node is represented by an arrow which describles the
R-rule subtended by the associated class;
• the height h (C) ∈ R+ of each node C satisfies the following condition: for
each node C 0 ∈ HV s.t. C 0 ⊂C
b then h (C) > h (C 0 ).

6.3 Construction of an implicative hierarchy

The significant R-rules which form an implicative hierarchy are calculated by

a incremental algorithm similar to the basic process of the classical HC. The
amalgamation criterium is here the maximization of the cohesion.
At each level hi of HV , a new R-rule is built. It results from the amalga-
mation of two R-rules built at a previous level hj , 0 < j < i. More precisely,
• the initial level h0 of HV is composed of the variable set V ;
• at h1 , two variables of V with the maximal cohesion are “grouped” together
to form a R-rule of degree 1;
• at h2 , the R-rule is composed either of two variables not yet aggregated,
called separate variables, or of the R-rule of degree 1 built at h1 and a sep-
arate variable. The selected R-rule is the one with the maximal cohesion;
• at h3 , the R-rule may be of three types: a R-rule of degree 1 composed
of two separate variables, a R-rule of degree 2 composed of a R-rule of
degree 1 built at h1 or h2 and a separate variable, or a R-rule of degree 3
composed of the two R-rules of degree 1 built at h1 and h2 .
• and so on. The process stops as soon each cohesion of the new potential
R-rules is null.
For instance, for the implicative hierarchy of the figure 1, the process stops
at h3 if the cohesion is null for the R-rules (a → (d → e)) → (b → c) and
(b → c) → (a → (e → d)).
We refer to [15] for an algorithmic description of this algorithm and the
analysis of its complexity.
The directed hierarchy HV can be associated with a valuation which sat-
isfies the ultrametric inequality.
 An overview of the Statistical Implicative Analysis (SIA) development 29

Proposition 5. [15] For any class C of HV , let us define the height h of C

by h (C) = 1 − c (C) if C is non elementary and h (C) = 0 otherwise, where
c (C) is the cohesion of the R-rule associated with C. Let u be a dissimilarity
on V × V defined by
• u (a, b) = 1 if a and b are not amalgameted in HV ,
• u (a, b) = h (Cab ) otherwise
where Cab is the smallest class of HV which contains both a and b. The dis-
similarity u is symmetric, positive and satisfies the ultrametric inequality:

u(a, b) ≤ Max {u (a, c) , u (b, c)}

for any a, b, c ∈ V .

From the Benzécri-Johnson theorem [4, 22] this property a posteriory jus-
tifies our choice of the word “hierarchy”.

7 The significative levels of the implicative hierarchy

7.1 The basic situation

Due to the multiplicity of the levels in the implicative hierarchy, it is nec-

essary to highlight those which are the more relevant for the structuration
process. In psycho-didactical or sociological applications, these levels seem to
correspond to consistent and stable conceptions. Hence, they contribute to a
finest interpretation of the set of the computed R-rules.
We have investigated two different approaches for this problem. The first
one is based on a rank analysis used in HC by Lerman [26]: it compares the
quality of the partitions obtained at each level of the hierarchy. The second
one is more local [19]: it focusses on the quality of the R-rules built at each
level. In the following, we present the first approach which is the only one to
be implemented in the CHIC software [9].

7.2 A criterium to determine the significative levels

Let us note that the cohesion coefficient defined in the section 5.3 can be
associated with a pre-ordering c on P = V × V − {(a, a) , (b, b) , . . .}:

(a, b) c (c, d) ⇔ c (a, b) ≤ c (c, d)

The idea consists in determining the levels of HV which “better express”

this pre-ordering. At each level hk , two sets of variable pairs can be distin-
guished: the set Ak of the amalgameted variable pairs at hk , and the set
Sk of the separate variable pairs (not yet amalgameted to form a R-rule of
degree ≥ 1). By construction, Ak ∪ Sk = P .
30 R. Gras, P. Kuntz

Let Gc be the graph of c . The set Gc ∩ (Ak × Sk ) is composed of

pairs of pairs which respect c at the level k. For instance, let us consider the
variable set V = {a, b, e, f } such that c (a, b) < c (e, f ). Let us suppose that at
the level hk the variables e, f and k are separate whereas the variables a and b
are amalgameted in a class. Then, the pair ((e, f ) , (a, b)) ∈ Gc ∩ (Ak × Sk ).
The objective is now to measure the adequation between Gc and Ak × Sk .
Let us denote by Θ the set of all the pre-orderings on P = V × V −
{(a, a) , (b, b) , . . .} with the same cardinality as c . We consider the random
preordering G∗ on Θ -with a uniform distribution-. From the theorem of Wald
and Wolfowitz [31] we can deduce that the theoretical mean of G∗ ∩ (Ak × Sk )
is µ = 1/2 card (Ak × Sk ) and its standard deviation is
1
σ= (card (Ak × Sk ) (cardG∗ + 1))
12
The adequation between Gc and Ak × Sk at the level hk is measured by

card (Gc ∩ (Ak × Sk )) − µ

s (c , k) =
σ

Definition 10. A level hk of the implicative hierarchy HV is significative if

it is a local maximum of s (c k): s (c , k − 1) < s (c , k) < s (c , k + 1)
If Gc ∩ (Ak × Sk ) = Ak × Sk then the partition Ak × Sk on V × V
associated with the structuration at hk is in total accordance with the pre-
ordering induced by the cohesion.

8 Typicality and contributions

8.1 The basic situation

Like in factorial analysis, we introduce the notion of “additional variable”: it

does not contribute to the computation of the relationships involved in the
implicative hierarchy, but it brings an additional information for its interpre-
tation(e.g. age, sex, social-professional category).
Our objective is to identify individuals, or individual groups, and addi-
tional variables which contribute to class forming at each level of the implica-
tive hierarchy.

8.2 A representation space

Let C be the class built at the level hk of the hierarchy HV . This class results
from the amalgamation of two classes C 0 ∈ HV and C 00 ∈ Hv not amalgameted
at the previous level hk−1 .
 An overview of the Statistical Implicative Analysis (SIA) development 31

The variable pair (a, b) is a generic pair at hk if ϕ (a, b) ≥ ϕ (i, j) for any
i ∈ C 0 and j ∈ C 00 . The generic intensity at hk is denoted by ϕk = ϕ (a, b).
This pair characterizes the most noticeable implicative effect for a given class.
Moreover, the classes C 0 and C 00 are themselves the results of an amalga-
mation at a lower level. Hence, at each level hg , g ≤ k, of HV , we can deter-
k
mine a generic pair: the resulting vector (ϕ1 , ϕ2 , . . . , ϕk ) ∈ [0, 1] is called the
implicative vector of the class C built at hk .
A similar representation can be used for evaluating the impact of an in-
dividual on the formation of a path on the implicative graph GM,α . Let us
consider a path P of length k on GM,α with a transitive closure (i.e. each
arc is associated with a rule with an implication intensity greater than 0.5).
Then, P contains k (k − 1) /2 transitive arcs. A pair (a, b) of P is generic if
ϕ (a, b) ≥ ϕ (i, j) for any i, j ∈ P .
k
The vectors (ϕ1 , ϕ2 , . . . , ϕk ) ∈ [0, 1] form a representation space where
the individuals can be projected. In the following, we precise the properties
of this space for an implicative hierarchy. They could be similarly defined for
an implicative graph.

8.3 Implicative power of an individual on a class

In this subsection, we define a dissimilarity on E × HV to measure the “prox-

imity” between an individual i ∈ E and a class C ∈ HV .
We first check if the individual i is in accordance with the implication of
the generic pair (a, b) of C at the level hk . Let us denote by a (i) (resp. b (i))
the binary variable which characterizes the presence/absence of a (resp. b) for
i. The contribution of i to the pair (a, b) is defined by
• ϕi,k = 1 if a(i) = 1 or 0 and b (i) = 1
• ϕi,k = 0 if a (i) = 1 and b (i) = 0
• ϕi,k = p ∈ ]0, 1[ if a (i) = b (i) = 0
In practice, p is set to the neutral value 0.5.
Any

individual i is associated with a k-dimensional vector ϕi,1 , ϕi,2 , . . . ,
ϕi,k which characterizes its contribution to the k generic pairs of the class
C buit at hk . An individual whose components are equal to the implicative
vector (ϕ1 , ϕ2 , . . . , ϕk ) is called the optimal typical individual.
We measure the typicality of i in C by the χ2 distance between the dis-
tributions (1 − ϕg ) and (1 − ϕi,g ), for g = 1, k. In contrast with the usual
Euclidean distance, it allows to compare ϕg − ϕi,g to ϕg and to normalize the
distance effect for large ϕg .

Definition 11. The implicative distance d2 (i, C) between an individual i ∈ E

and a class C ∈ HV built at the level hk is defined by
32 R. Gras, P. Kuntz

k
!1/2
2
1 X (ϕg − ϕi,g )
d1 (i, C) =
k g=1 1 − ϕg

If it exists g s.t. ϕg = 0 we set (ϕg − ϕi,g ) / (1 − ϕg ) = 0. In this case,

the generic implication is maximal and thus it exists an excellent implicative
relationship for all the individuals i ∈ E (ϕi,g = 1).

Remark 2. Let us consider a class C ∈ HV at the level hk . We can define a

metric space structure on E with

k
!1/2
2
1 X (ϕi,g − ϕj,g )
dC (i, j) =
k g=1 1 − ϕg

for any (i, j) ∈ E 2 .

The distance dC (i, j) measures the behavior difference between i and j consid-
ering C. It defines a discrete topological C-structure
− on E. Let us consider the
→ − →
vectors (ϕi,1 , ϕi,2 , . . . , ϕi,k ) and the norm i − j = dC (i, j). This topology

is equivalent to the previous one (similarly to the duality in correspondence

analysis). The elements of the diagonal matrix of the symmetrical operator
−1
associated with the quadratic form which defines dC are (k (1 − ϕi )) for
i = 1, k. Let us remark that the semantic of the vector sum is not precised in
the SIA. Nevertheless, it could be interesting to characterize the individuals
which belong to a ball of a given diameter with a given center (e.g. the optimal
individual).

8.4 Individual and group typicalities

Definition 12. The typicality γ (i, C) of an individual i ∈ E for a class

C ∈ HV is defined by the ratio between the distance d1 (i, C) and the maximal
value of the distance on the individual set:

d1 (i, C)
γ (i, C) = 1 −
Maxj∈E d1 (j, C)

The maximal distance d1 (j, C) is reached by the individuals with null or

very low ϕi,k . They are contrasting with the generic rules. And, the typicality
of i is large when i is different from these individuals.
A straightforward extension of the previous definition allows to define the
typicality γ (G, C) of a individual group G ⊂ E:
1 X
γ (G, C) = γ (i, C)
card (G)
i∈G
 An overview of the Statistical Implicative Analysis (SIA) development 33

In practice, an operational tool is required to evaluate the statistical sig-

nificancy of a group typicality. The basic idea consists in partitioning E in
two opposite groups E1 and E2 with regards to their typicalities γ (E1 , C) and
γ (E2 , C) in C. This dispersion can be measured by the inter-class inertia. The
barycenter γ of the typicalities γ (E1 , C) and γ (E2 , C) is defined by
1
γ= (card (E1 ) γ (E1 , C) + card (E2 ) γ (E2 , C))
n
By construction, γ is also the barycenter of all the individual typicalities
in E. Consequently, the inter-class inertia is
card (E1 ) 2 card (E2 ) 2
VE = (γ (E1 , C) − γ) + (γ (E2 , C) − γ)
n n

Definition 13. An individual group G∗C ⊂ E is optimal for a class C ∈ HV

if its typicality is greater than the typicality of its complementary set in E,
and if it constitutes with this later a bi-partitioning which maximizes VE . This
partition is said to be significant.
It is interesting to detect the group or the additional variable associated
with the greatest typicality for the optimal group. We measure the surprising-
ness of the proportion of concerned individuals. Let {Ei }i be a given partition
of E. It can be defined by an additional variable. For each class Ei , we con-
sider the random variable Xi which is a random subset of E of cardinality
card (Ei ), and the random variable Zi defined by Zi = card (Ei ∩ G∗C ). The
variable Zi follows a Binomial distribution with parameters card (Ei ) and
card (G∗C ) /n [21].

Definition 14. The most typical group of the class C is the subset Ei ⊂ E
which minimizes the probability pi on the set {pi = Pr (Zi > card (Xi ∩ G∗C ))}i .
The probability pi is an error of the first kind: the risk of making a mistake
when considering that the group is not typical.

8.5 Contribution

The contribution is different from the typicality: it measures the individual

and additional variable responsabilities for the existence of a rule or a R-rule
between the variables of V .
Let us consider two variables a ∈ V and b ∈ V linked by a rule a → b
at the first level h1 of the implicative hierarchy Hv . The contribution of an
individual i to (a, b) is defined by ϕi,1 . This notion can be extended to the
formation of a class C at the level hk .
34 R. Gras, P. Kuntz

Definition 15. The distance d2 (i, C) between an individual i ∈ E and a class

C at the level hk of the implicative hierarchy HV is defined by
k
1X 2
d2 (i, C) = (1 − ϕi,g )
k g=1

The contribution θ (i, C) of i ∈ E to C ∈ HV is defined by θ (i, C) =

1 − d2 (i, C).

The maximal value of θ (i, C) is equal to 1; it is reached for an individual

i whose components ϕi,g are all equal to 1.
The concepts defined in the previous sections can be easily adapted to the
distance d2 . In practice, the contribution is often easier to interpret than the
typicality.

9 Illustration
We illustrate the applicative interest of the different concepts presented be-
low on a data set stemming from a survey of the French Public Education
Mathematical Teacher Society on the level in mathematics of pupils in the
final year of secondary education and the perception of this subject [8]. In
parallel with evaluation tests for students, a set of 311 teachers have been
asked on the objectives of the training in mathematics (table 2 presents some
items used in the following) and their opinions about commonly shared ideas
on this subject (table 3). For each proposition, the teacher could answer
“I agree with this idea” (positive opinion), “I disagree” (negative opinion) or
“I partially agree”.
The figure 3 presents a part of the directed hierarchy obtained on the set
composed of the objectives and the different modalities for the opinions (51
items). The interpretation of the whole set of rules is far beyond the scope of
this paper. Nevertheless, we have selected some of them, easy to interpret for
a non specialist in education theory, to show the use of a directed hierarchy
on a real-life corpus. As for the complementarity of this structure with a more
classical approach based on the relationship representation by a graph, it is
highlighted in figure 2. The vertex set V of this graph contains the same items
as those selected for figure, and there is an arc between two vertices ai and
aj of V if and only if ϕ (ai , aj ) ≥ 0.5 and for any ak ∈ V , ϕ (ai , ak ) < 0.5 and
ϕ (aj , ak ) < 0.5 (e.g. [2]). The choice of the threshold comes from the fact that
beyond 0.5 the implicative tendency (e.g. ai → aj ) is better than neutrality.
It is important to note that, due to the non transitivity of the relationship
on A induced by ϕ, the existence of two arcs of the form (ai , aj ) and (aj , ak )
does not entail the existence of the arc (ai , ak ). For instance, in figure, we can
not deduce a relationship between the items E and OP 7.
 An overview of the Statistical Implicative Analysis (SIA) development 35

E OP2

I OP5 N

OP8 OP4 A

OP7 OP6

Fig. 2. A part of the implicative graph on the items of the survey on the training
in mathematics

Fig. 3. A part of the directed hierarchy on the items of the survey on the training
in mathematics
36 R. Gras, P. Kuntz

On the other hand, beside the binary rules, most of the R-rules of the total
directed hierarchy involve three or four items. The interpretation of rules with
more attributes are generally more difficult to interpret. Nevertheless, they
provide more information than the set of the implied binary rules.
The R-rule (N → A) → OP 6 has the following meaning: if know-how
acquisition must be accompanied by knowledge acquisition, then the teacher
ask for well-defined programs. In this case, focussing on knowledge requires
a predefined charter from the institution. The R-rule allows to give a more
synthetical interpretation than the binary rules: these are concerned with
the behaviour, as seen within the behavioural framework, whereas the R-rule
here describes a conduct of a higher order which determines the behaviour.
Teachers who consider that the objective C (Preparation to civic and social
life) is not relevant are mostly responsible for this R-rule. They have a very
restrictive representation of the teaching of maths, focussed on the subject,
and their teaching conforms to national standard without any questioning.
The R-rule (OP 2 → (OP 5 → OP 4)) can be interpreted as follows: if I wish
to keep up the complete problem for the A-level exam and if the importance
given to the demonstration in maths is subordinated to a fixed scale of grading,
then I conform to the national syllabus instructions. This rule corresponds
to a class of teachers subjected to the institution and conservative in their
educational choices. They consider that, in France, the land of Descartes,
the demonstration is the foundation of the mathematical activity and that
the complete problem at the exam is the evaluation criterion. For them, the
syllabuses and the grading scales defined by the institution are essential to
teaching and assessment. We find again a very classical teaching conception
based on an explicit and unconditional support to the institution.
Contrary to the previous ones, the R-rule (I → (E → (OP 8 → OP 7)))
can be interpreted as a sign of an openminded didactic conception. Indeed, it
means that if a teacher lays the emphasis on the critical mind development
and the imagination and creativity, then he considers that a personal train-
ing of the pupils in the search of examples and counter-examples is sufficient
for discovering divisibility features by themselves. This R-rule reveals a rela-
tionship between the non-dogmatic behaviours of the teacher and the wish to
place the pupil in a situation of personal research.

A Knowledge acquisition
B Preparation to professional life
C Preparation to civic and social life
D Preparation to examinations
E Development of imagination and creativity
I Development of critical mind
N Know-how acquisition

Table 2. Some items from the list of the objectives of training in mathematics
 An overview of the Statistical Implicative Analysis (SIA) development 37

OP1 It’s true that maths are an element of selection

OP2 For the A-level exam, I prefer a complete problem with different
parts rather than independent questions
OP3 In my grading system, I give more importance to the reasoning
than to the result
OP3 When I correct, I prefer a very detailed grading system
OP5 The demonstration is the only rigourous way to do maths
OP6 I prefer well-defined programs precising what I must do and not do
OP7 In the last form of secondary education, a pupil should be able
to recognize whether a number written in the base 10 is divisible by 4
OP8 In the last form of secondary education, a pupil should be able to give
an example or counter-example of the following statement: if two
applications f and g are strictly increasing on a given interval, then
the product f × g is also increasing.
OPX Individual estimation of a size (e.g. width, length)

Table 3. Some items from the list of the commonly shared ideas in the teaching of
maths

We study now the additional information brought by the supplementary

variable which defines the main option of the cursus: Scientific (S), Economic
and Social (ES), Arts (A) and Technology (T ). The observed distribution of
the variable is: S = 155, ES = 68, A = 22 and T = 66.
Let us consider the class C = (E → (OP 8 → OP 7)) → OP X. This rule
corresponds to a class of teachers which give importance to imagination and
personnal research. The most typical modality for this variable is S (scientific).
Indeed, 116 teachers of the option S on 155 are in the optimal group G∗C of
cardinality 201. Let X be a random subset of same cardinality as S (155) and
Z be random variable defined by the intersection of X and the optimal group
G∗ . Then, Z follows a binomial distribution of parameters 155 and 201/311 =
0.656. The probability for Z to be greater than 116 is the risk 0.00393. The
analysis of the series of the risks associated with the different options S, A
and T shows that the most typical modality of the class C is S. The pair
(S, C) is said to be mutually specific. Similarly, the most typical modality of
the rule B → K is T ; consequently, the pair (T, (B, K)) is mutually specific. It
confirms that the teachers in technical cursus consider that the mathematics
should be useful for the professional life (B) and consequently for the other
disciplines.
The computation of the contribution of S to C shows that 111 teachers
on 311 participate to the optimal group. The number of teachers of S has
decreased (from 116 to 67), and its proportion in the optimal group is signif-
icantly lower than for the typicality computation. The teachers of S are the
most typical, i.e. in accordance with the general behavior of the population.
However, their contribution to the four involved variables is lower than the
contribution of the teachers of the other cursus. The risk is equal to 0.0251: it
38 R. Gras, P. Kuntz

is more than 6 times greater than the typicality. This remark illustrates the
nuances brought by the two concepts: typicality and contribution.

10 Conclusion
In this paper we have proposed an overview of the Statistical Implicative
Analysis. Beyond the results, we have related the genesis of the considered
problems which arise from questions of experts in different fields. The the-
oretical basis is quite simple, but the numerous questions on the original
assumptions, which do not appear here, have lead to modifications and some-
times to deeper revisions. Fortunately, the proposed answers go beyond the
original framework, and SIA is now a data analysis method, based on a non
symmetrical approach, which has been shown to be relevant for various ap-
plications.
In the next future, we are planning to consider new problems: (i) the exten-
sion of SIA to vectorial data, (ii) and to fuzzy variables, (iii) the integration
of missing data, (iv) the redundant rule reduction. We are also interested in
the complementarity of SIA with other approaches, in particular with decision
trees (see Ritschard’s paper in this book). And, we will obviously carry on ex-
ploring real-life data sets and confronting our theoretical tools to experimental
analysis to make them evolve.

References
1. R. Agrawal, T. Imielinsky, and A. Swami. Mining association rules between sets
of items in large databases. In Proc. of the ACM SIGMOD’93, pages 679–696.
AAAI Press, 1993.
2. M. Bailleul. Des réseaux implicatifs pour mettre en évidence des relations.
Mathématiques, Informatique et Sciences Humaines, 154:31–46, 2001.
3. M. Bailleul and R. Gras. L’implication statistique entre variables modales.
Mathématiques et Sciences Humaines, 128:41–57, 1995.
4. J.P Benzécri. L’analyse des données (vol. 1): Taxonomie. Dunod, Paris, 1973.
5. J.M. Bernard and S. Poitrenaud. L’analyse implicative bayesienne d’un ques-
tionnaire binaire : quasi-implications et treillis de galois simplifié. Mathéma-
tiques, Informatique et Sciences Humaines, 147:25–46, 1999.
6. J. Blanchard, P. Kuntz, F. Guillet, and R. Gras. Mesure de la qualité des
règles d’association par l’intensité entropique. Revue des Nouvelles Technologies
de l’Information-Numéro spécial Mesures de qualité pour la fouille de données,
RNTI-E-1:33–44, 2004.
7. J. Blanchard, P. Kuntz, G. Guillet, and R. Gras. Implication intensity: From the
basic definition to the entropic version - chapter 28. In Statistical Data Mining
and Knowledge Discovery, pages 475–493. CRC Press - Chapman et al., 2003.
8. A. Bodin and R. Gras. Analyse du préquestionnaire enseignants. Bulletin
de l’Association des Professeurs de Mathématiques de l’Enseignement PUblic,
425:772–786, 1999.
 An overview of the Statistical Implicative Analysis (SIA) development 39

9. R. Couturier and R. Gras. C.h.i.c. : Traitement de données avec l’analyse im-

plicative. Revue des Nouvelles Technologies de l’Information, RNTI-II:679–684,
2005.
10. L. Fleury. Extraction de connaissances dans une base de données pour la gestion
de ressources humaines. PhD thesis, Université de Nantes, 1996.
11. R. Gras. Contribution à l’étude expérimentale et à l’analyse de certaines acqui-
sitions cognitives et de certains objectifs en didactique des mathématiques. PhD
thesis, Université de Rennes 1, 1979.
12. R. Gras, S. Ag Almouloud, M. Bailleul, A. Larher, M. Polo, H. Ratsimba-
Rajohn, and A. Totohasina. L’implication statistique - Nouvelle méthode ex-
ploratoire de données. La Pensee Sauvage editions, France, 1996.
13. R. Gras, R. Couturier, J. Blanchard, H. Briand, P. Kuntz, and P. Peter.
Quelques critères pour une mesure de qualité de règles d’association. Revue
des Nouvelles Technologies de l’Information, RNTI-E-1:197–202, 2004.
14. R. Gras, E. Diday, P. Kuntz, and R. Couturier. Variables sur intervalles et
variables-intervalles en analyse statistique implicative. In Proc. of Société Fran-
cophone de Classification, pages 166–173. Université des Antilles-Guyane, 2001.
15. R. Gras and P. Kuntz. Discovering r-rules with a directed hierarchy. Soft
computing, 1:46–58, 2005.
16. R. Gras, P. Kuntz, and H. Briand. Les fondements de l’analyse statistique
implicative et quelques prolongements pour la fouille de données. Mathématiques
et Sciences Humaines, 154:9–29, 2001.
17. R. Gras, P. Kuntz, and H. Briand. Hiérarchie orientée de régles généralisées en
analyse implicative. Extraction des Connaissances et Apprentissage, 17-3:145–
157, 2003.
18. R. Gras, P. Kuntz, R. Couturier, and F. Guillet. Une version entropique de
l’intensité d’implication pour les corpus volumineux. Extraction des Connais-
sances et Apprentissage, 1-2:69–80, 2001.
19. R. Gras, P. Kuntz, and J.-C. Régnier. Significativité des niveaux d’une hiérarchie
orientée en analyse statistique implicative. Revue des Nouvelles Technologies de
l’Information, RNTI-C-1:39–50, 2004.
20. R. Gras and A. Larher. L’implication statistique, une nouvelle méthode
d’analyse de données. Mathématiques, Informatique et Sciences Humaines,
120:5–31, 1992.
21. R. Gras and H. Ratsimba-Rajohn. Analyse non symétrique de données par
l’implication statistique. RAIRO-Recherche opérationnelle, 30-3:217–232, 1996.
22. S.C. Johnson. Hierarchical clustering scheme. Psychometrika, 32:241–254, 1967.
23. P. Kuntz, R. Gras, and J. Blanchard. Discovering extended rules with implica-
tive hierarchies. In Proc. of the new frontiers of statistical data mining and
knowledge discovery, pages 166–173. Knoxville, Tennesee, 2001.
24. J.B. Lagrange. Analyse implicative d’un ensemble de variables numériques:
application au traitement d’un questionnaire aux réponses modales ordonnées.
Revue de statistique appliquée, 46(1):71–93, 1998.
25. P. Lenca, P. Meyer, B. Vaillant, P. Picouet, and S. Lallich. Evaluation et analyse
multi-critères des mesures de qualité des règles d’association. Revue des Nou-
velles Technologies de l’Information, RNTI-E-1:219–246, 2004.
26. I.C. Lerman. Classification et analyse ordinale des données. Dunod, Paris, 1981.
27. J. Loevinger. A systemic approach to the construcion and evalation of tests of
ability. Psychological Monographs, 61, 1947.
40 R. Gras, P. Kuntz

28. J. Piaget. Le jugement et le raisonnement chez l’enfant. Delachaux et Niestlé,

1967.
29. J.-C. Régnier and R. Gras. Statistique de rangs et analyse statistique implica-
tive. Revue de Statistique Appliquée, LIII:5–38, 2005.
30. L. Seve. Emergence, complexité et dialectique. Odile Jacob, Paris, 2005.
31. A. Wald and J. Wolfowitz. Statistical tests based on permutations of the obser-
vations. Ann. Math. Stat., 15, 1944.
CHIC: Cohesive Hierarchical Implicative
Classification

Raphaël Couturier

Computer Science Laboratory of University of Franche-Comte (LIFC),

IUT de Belfort-Montbeliard, BP 527, 90016 Belfort, France
[email protected]

Summary. CHIC is a data analysis tool based on SIA. Its aim is to discover the
more relevant implications between states of different variables. It proposes two
different ways to organize these implications into systems: i) In the form of an
oriented hierarchical tree and ii) as an implication graph. Besides, it also produces
a (non oriented) similarity tree based on the likelihood of the links between states.
The paper describes its main features and its usage.

Key words: data mining tool, oriented hierarchical tree, implication graph, simi-
larity tree, CHIC.

1 Introduction
Statistical Implicative Analysis was initiated by Gras [7, 8]. The first goal of
this method was to define a way of answering the question: “If an object has
a property, does it also have another one? ”. Of course the answer is rarely
true. Nevertheless it is possible to notice that a trend is appearing. SIA aims
at highlighting such tendencies in a set of properties. SIA can be considered
as a method to produce association rules. Compared to other association rule
methods, SIA distinguishes itself by providing a non linear measure that sat-
isfies some important criteria. First of all, the method is based on implication
intensity that measures the degree of astonishment inherent in a rule. Hence,
some trivial rules that are potentially well known to an expert are discarded.
In fact, a rule of the form A ⇒ B is considered trivial if almost all objects
of the population have property B. In this case, the implication intensity is
close to 0 and this is not the case when rules can be considered as surprising.
This implication intensity may be reinforced by the degree of validity that
is based on Shannon’s entropy, if the user chooses this computation mode.
This measure does not only take into account the validity of a rule itself, but
its counterpart too. Indeed, when an association rule is estimated as valid,
i.e. the set of items A is strongly associated with the set of items B, then

R. Couturier: CHIC: Cohesive Hierarchical Implicative Classification, Studies in Computational

Intelligence (SCI) 127, 41–53 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
42 Raphaël Couturier

it is legitimate and intuitive to expect that its counterpart is also valid, i.e.
the set of non-B items is strongly associated with the set of non-A items.
Both the implication intensity and the degree of validity can be completed
by a classical utility measure based on the size of the support rule and are
combined to define a final relevance measure that inherits the qualities of the
three measures (with the entropic theory), i.e. it is noise-resistant as the rule
counterpart is taken into account and it only selects non trivial rules. For
further information the reader is invited to consult [9]. Based on that original
measure, CHIC, given a set of data, enables one to extract association rules.
CHIC and SIA have been used in wide domain areas, for example [3, 4, 6, 14].
Based on the implication intensity and the similarity intensity CHIC allows
to build two trees and one graph. The most classical tree is a similarity tree
(usually known as dendogram). It is based on the similarity index defined by
Lerman [13]. In a similar way, the implication intensity can be used to build
an oriented hierarchy tree. The implication intensity can also be used to define
an implication graph, which lets the user select the association rules and the
variables he or she wants.
In opposition to most of other multidimensional data analysis methods,
the SIA establishes the following properties between the variables it handles:
• relationship between variables are dissymmetrical
• the association measures are non linear and are based on probabilities
• the user can use graphical representations which follow the semantic of the
relationship
For example, most of the following methods: the factor analysis, the discrim-
inant analysis or the preference analysis are based on metric space distances.
Most of hierarchical classification methods use proximity or similar indexes.
So, relationships between variables are essentially symmetric. Moreover, most
of the times those relationships vary linearly with observation parameters.
Some methods are built on measures based on probabilities which simplifies
the results interpretation. Some papers present comparison between different
measures. Interested readers can consult [12] or the chapter entitled “On the
behaviour of the generalisation of the intensity of implication : a data-driven
comparative study” in this book.
Section 2 addresses the variables that can be handled in CHIC, their format
and the options that may help users. In Section 3 some details are explained
on the way to compute the association rules. Section 4 presents the similarity
and the hierarchy tree. Section 5 describes the implication graph. In Section 6
some other features of CHIC are presented. Section 7 gives an illustration
with interval variables and computation of typicality and contribution. Finally
Section 8 concludes this paper.
 CHIC: Cohesive Hierarchical Implicative Classification 43

2 Variables

Initially CHIC as the SIA was designed to handle binary variables. Later,
SIA was enhanced by other kinds of variables and so was CHIC. Currently,
CHIC allows the user to handle binary variables, frequency variables, variables
over intervals and interval-variables. The case of binary variables is obviously
the simplest one. Ordinal variables (also called nominal ones) can be coded
using as many binary variables as number of categories. Frequency variables
take a real value between 0 and 1. This kind of variables allows the user
to include the case of discrete variables which only takes a fixed number of
values (or modalities) ranging between 0 and 1. Of course, the way of defining
modalities is very important, because it strongly affects the results of CHIC
whether the values of modalities are close to 0 or 1. This remark is also
true concerning the frequency variables. It should be noticed that ordinal
variables are also coded using frequency ones. The user must pay attention to
the way real variables are transformed into frequency ones. Several strategies
are available depending on the values. If the values are positive, they can
be divided by the maximum value. Another possibility resides in considering
that the minimum value represents 0 and the maximum represents 1, all the
other variables are proportionally distributed between the minimum and the
maximum values. If a real variable has both positive and negative values, it
is possible to split the variables into two variables, one for positive values
and another one for negative values. In this case, previous remarks are true
for both new variables. However, it is possible to consider that the minimum
value (even if it is negative) represents 0 and the maximum represents 1. In
this case, all other values are transformed into the interval [0, 1].
Variables over intervals and interval-variables are used to model more com-
plex situations. Both these kinds of variables are explained in the following
section. Variables over intervals allow to stage the following problem. In fact,
the conversion of a real variable into a frequency one may imply difficult
choices from the user’s point of view, as explained previously. Using the same
real values, a variable over intervals proceeds differently. It consists in decom-
posing values of a variable into a given number of intervals. The number of
intervals is chosen by the user and then the algorithm of dynamic clouds [5]
automatically constitutes the intervals which have distinct bounds. This al-
gorithm has the particularity of building intervals by minimizing inertia in
each interval. Then each interval is represented by a binary variable and an
individual has value 1 if it belongs to the interval and 0 otherwise. Using such
a decomposition, an individual belongs only to one interval. Hence, the num-
ber of variables increases with this method. Let us take an example. Assume
that we have a set of individuals and that for each of them we have their
weight and height. Then, assume that everybody weighs between 40kg and
140kg and the height ranges between 140cm and 200cm. Figure 1 shows an
example with few individuals, values have been chosen arbitrarily. Supposing
that we are interested in decomposing each variable into 4 intervals, according
44 Raphaël Couturier

Fig. 1. A simple example of data with interval variables and supplementary variables

to the distribution of both variables, it is possible that we obtain the follow-

ing intervals [40, 60[, [60, 95[, [95, 110[, [110, 140] for the weights which are
respectively called weight1, weight2, weight3 and weight4 and the following
intervals [140, 165[, [165, 174[, [174, 186[, [186, 200] for the heights which are
respectively called height1, height2, height3 and height4. In the follow-up to
this computation, all the unions of the intervals of a variable are considered.
So with variable height, there are also intervals height12, height23, height34,
height1−3, height2−4, height1−4. Intervals of the form nameAB correspond
to the union of two consecutive intervals, for example height23 corresponds to
the union of height2 and height3. Intervals of the form nameA−B correspond
to the union of all the consecutive interval between nameA and nameB, for
example height1−3 corresponds to the union of height1, height2 and height3.
Of course the most interesting feature of the interval variables consists in try-
ing to make smaller partitions with these intervals, i.e. merging some intervals
together in order to know which intervals are naturally close to each other.
CHIC implements such an algorithm which is mathematically described, for
example, in [7,8]. With the previous example, if other variables inform on the
habits of individuals, it is then possible to obtain information about possible
relationships between these other variables and the weights and the heights
of people. For example, it is possible to know that people measuring between
140cm and 180cm are best suited to doing some particular things or that
other people with such and such habits weigh principally between 90kg and
 CHIC: Cohesive Hierarchical Implicative Classification 45

150kg. Of course the number of intervals may have a great influence on the
result.
Whereas for a variable over intervals each individual takes a value 1 for
only one interval, the particularity of an interval-variable is that an individ-
ual takes some values on different intervals. Moreover the intervals can be
contiguous and represent a discrete decomposition, as it is the case using an
automatic decomposition method like the dynamic clouds one, but they can
also be defined by the user according to appropriate criteria. Taking the pre-
vious example with the weight and height, a user may prefer to state that
people may be thin, normal or healthy and that they can be small, normal
or tall. Nonetheless, the particularity of an interval-variable is that an indi-
vidual may take values between several intervals but the sum of all its values
must be less or equal to 1. In most cases, the sum will be equal to 1, but
this is not mandatory. Roughly speaking, it is far from being easy to classify
objects and individuals because opinions may frequently diverge on the fact
that somebody or something should be described as “small” or “normal” for
example. Consequently, saying that someone is rather slim may be expressed
by assigning this individual with 0.75 thin and 0.25 normal. It should also
be noted that this allows the user to handle fuzzy variables which are very
useful in several problems [2]. The fuzziness characteristic comes from either
a human appreciation, which by definition is subjective, or by an inaccurate
measurement process which for some reason introduces a bias. In any case,
CHIC uses the standard methods presented in the next section.
As to the data format, CHIC uses the CSV (comma separated values)
format, a standard in spreadsheet tools. Labels for individuals are recorded
on the first row and labels for variables are recorded on the first column.
Values of individuals are represented into a 2-dimensional array. The values
for each variable of an individual are stored into a line in this array (the
first element is the name of the individual). The values for each individual
of a variable are stored into a column in this array (the first element is the
name of the variable). Of course, the nature of the values in the array differs
according to the kind of variables (binary, frequency variable, . . . ).
As explained in the article of Gras and Kuntz in this book, supplementary
variables can be used in CHIC in order to explain some important facts.
This kind of variables do not intervene in the computation but it is used to
give sense for the computation of typicality and contribution. Let us take an
example. Assuming that we want to study the impact of a new tramway in
a part of a town and that in this aim a survey has been performed. This
survey gathers several information concerning the needs and the hopes of
this project’s potential users. Of course the gender of the people questioned
is given. For example, some rules such as: working people living far from
their work are generally very interested in the project, or family with young
children are also very favorable to it, may be extracted. Using the gender of
people as a supplementary variable, it is then possible to know if people that
46 Raphaël Couturier

are responsible for the construction of the previous rules are rather men or
women or if there is no distinction.
Before starting any computation with an appropriate presentation of the
results, the user must choose some options. The most important one is choos-
ing the computation method: either the classical one or the entropic one. This
criterion will produce different results. The entropic version of the implication
does not only take into account the validity of a rule itself, but its counterpart
too. Usually, the entropic version is best suited for large data set. It is also
more severe than the classical one which produces more intensive rules but is
totally inappropriate with a large data set.

3 How to efficiently compute association rules

At the start of a computation, CHIC first computes rules by choosing a sim-
ilarity analysis or an implicative one. The computation of association rules
(for implicative analysis) is based on the algorithm described by Agrawal [1].
This algorithm can efficiently compute conjunction rules. Roughly speaking,
in order to produce rules with n variables, it consists in computing all the
occurrences of all the possible tuples of variables of size 1 up to n. For ex-
ample, assuming that we have 5 variables labeled from A to E and that we
are interested in finding rules composed of 3 variables, i.e. rules of the form
A ∧ B ⇒ C, then the algorithm seeks frequent co-occurrences of all the vari-
ables, of all the pairs of variables and of the triplets of variables. With the
example, the triplets are: ABC, ABD, ABE, ACD, ACE, ADE, BCD, BCE
and CDE. If the user chooses a given threshold α and assuming that |B| < α
(where |B| represents the cardinality of B) and all other variables have a
larger cardinality, then only 4 triplets may have a cardinality greater than α
(ACD, ACE, ADE and CDE), but this is not necessarily the case. Then for
all the tuples whose occurrences are greater than the threshold, it is possible
to produce the rules. For example with the following tuples ABC, AB and
BC, AC, it is possible to compute the intensity of the two rules A ∧ B ⇒ C,
B ∧ C ⇒ A and A ∧ C ⇒ B.
Even if the user specifies a high threshold, the number of rules produced by
the algorithm with large data sets may be quite huge. Moreover, it is possible
that some rules are very similar, that is why we have introduced an originality
criterion. It allows one to select only conjunction rules that are original, i.e.
rules for which sub-rules are not so trivial. Consider, for example, the following
rule A ∧ B ⇒ C, this rule is original if its intensity is high and if the rules
A ⇒ C and B ⇒ C have a small intensity measure.
As the definition of the originality has never been described formally we
will explain it in details. Let Am ⇒ B be an association rule such that the
antecedent Am is a set of m properties, a set of properties is also viewed as
a conjunction. As soon as a rule Ai ⇒ B, where Ai is a subset of Am , has a
high implication intensity then, the originality of Am ⇒ B decreases.
 CHIC: Cohesive Hierarchical Implicative Classification 47

Therefore, the originality of Am ⇒ B depends on the implication intensity

associated with each rule Ai ⇒ B, i ∈ {1, . . . , m − 1}, such that Ai is a subset

of Am that contains i properties. For each i ∈ {1, . . . , m − 1} there is mi
i,j
subsets of Am that contains i properties.
We note Am the j − th subset with
i ∈ {1, . . . , m − 1} and j ∈ {1, . . . , mi }.
Thus, the originality of Am ⇒ B depends on the 2m −2 parts of Am (empty
set and Am are not considered). That is to say, the implication intensity of
Am ⇒ B
depends on the set of rules Rm = ∪i∈{1,...,m−1} {Ai,j m ⇒ B | j ∈
m
{1, . . . , i }}.
We propose to measure the originality of the rule Am ⇒ B as the geo-
metrical average of the difference of the implication intensity associated with
Am ⇒ B and with each rule of Rm . Finally, the coefficient of originality of
Am ⇒ B is defined as follow.
The originality of an association rule Am ⇒ B, where Am is a conjunction
of m properties, is measured by:
Originality(Am ⇒ B) =
1
    m
m 2 −2
 m−1 i 
 Y Y 
( ( ImpInt(Am ⇒ B) − ImpInt(Ai,j
m ⇒ B)))
 
 
 i=1 j=1 

Where ImpInt(A ⇒ B) is the implication intensity of the rule A ⇒ B.

With this definition, the number of “original” rules decreases often drasti-
cally. In general, when the number of conjunctions increases, the number of
trivial rules also increases. That is why the criterion of originality provides an
efficient way of only having interesting rules. We plan to compare our original-
ity measure with other measures that could produce similar results in future
works.

4 Similarity and hierarchy tree

Once CHIC has computed the whole set of rules according to the parameters
that the user has chosen, it can build a tree with some of the rules. This
tree may be seen as a classification oriented or not in function of the kind of
computation (similarity or implication). There are some common principles in
building both trees. In the following a rule is called a class and it is composed
of two variables in their simplest form. At each level of the classification,
CHIC selects the class with the highest intensity (of similarity or implication).
In order to know how to compute the intensity of a class and to have an
explanation of the variables used in figures 2, 3 and 4, one should refer to the
chapter of R. Gras and P. Kuntz in this book.
48 Raphaël Couturier

Fig. 2. An example of a similarity tree

Then, at each step, CHIC computes a set of new classes with all the ex-
isting ones. In order to build a new class, CHIC either aggregates an existing
class with a variable which has not been aggregated in another class yet or
aggregates two existing non aggregated classes. Nonetheless each couple of
variables between the two classes must have a valid intensity, i.e. greater than
0.5. For example the formation of a class ((a, b), c) entails that classes (a, c)
and (b, c) are meaningful from the analysis point of view (similarity or implica-
tion). The class ((a, b), c) represents the rule (a ⇒ b) ⇒ c with the implicative
analysis and represents the fact that a and b are similar and that this class
is similar to c from the similarity point of view. For more details on the class
formation, interested readers are invited to read the article of Gras and Kuntz
in this book and in [10].
If the user is interested to know how the tree without one or more variables
would look like, he can simply deselect them in the item toolbox. It should be
noticed that this toolbox is available for all kinds of representation provided by
CHIC (trees of graph). Unfortunately, a modification of the variables involved
in the computation (even a small one) implies a complete rebuilding of the
tree. This step of class construction is strongly dependent on the number of
variables (the algorithm has a complexity which depends on the factorial of
the number of variables).
Before running an analysis the user can choose in the computation options
to highlight the significant level in the tree.
Figure 2 shows a similarity tree and Figure 3 shows a hierarchy tree.
For the latter one, significant levels are pointed out. They are represented
by a red line (in CHIC). Each significant level means that the current level
 CHIC: Cohesive Hierarchical Implicative Classification 49

Fig. 3. An example of a hierarchy tree

is more significant than the previous one and than the next one, which is
not significant by definition. For more details on its construction, interested
readers should refer to the definition given in this book and in [11].
The similarity index is computed by the classical theory or by the entropic
one. The last one should be preferred with a large number of individuals.
Moreover, the construction of the similarity tree with the classical index leads
to only one class that gather all the others. On the contrary, with the entropic
version of the similarity index, the algorithm very frequently builds more than
one class. In fact, according to similarities of data, the number of classes varies.

5 Implication graph

As explained previously, both classifications (based on the similarity and im-

plication) in CHIC only select some of the rules and ignore some other rules by
constructing the tree. If all the rules are required to point out an interesting
feature, the implication graph may be preferred since in this graph, the user
can see the rules that have a greater intensity than a given threshold. In fact,
four thresholds are available and CHIC uses different colors to quickly show
which rules are the most important. In Figure 4 some rules are represented
in an implication graph. An arrow is used to show the implication between
two variables (the rule A ⇒ B is represented by an arrow from A to B).
As the number of rules may be large, the user has the possibility to select
only some variables. Hence, only rules with present variables are represented.
This consequently reduces the number of rules. Moreover, in order to make
the graph more readable, CHIC uses an automatic graph drawing algorithm
50 Raphaël Couturier

which tries to minimize the number of crossings between rules. By default,

transitive closures are not displayed on the implication graph. A simple click
with the mouse in the toolbox displays them. CHIC computes them once and
for all for each new graph. Afterwards, even if the user selects or deselects
some variables, changes the thresholds of the rules, chooses or not to display
transitive closures, then CHIC only displays the graph without any computa-
tion. This allows users to stress the important features of their data. Besides,
the graph drawing procedure may be time consuming with large graph, so it
is not used automatically.

Fig. 4. An example of an implication graph

6 Other possibilities

In addition to the traditional representation modes previously described,

CHIC provides some practical features. For each of the graphic representation
it is possible to compute the contribution and the typicality of an individual
to a given rule. In the same way, it is possible to compute the contribution
and the typicality of a set of individuals to a given rule.
The notion of typicality is defined by the fact that some individuals
are “typical” of the behavior of the population. They contribute well to the
creation of a rule, i.e. with a similar intensity to the rule. For example if a rule
 CHIC: Cohesive Hierarchical Implicative Classification 51

A ⇒ B has an implicative intensity equals to 0.7, then the most typical indi-
viduals respectively have values close to 0.5 and 1 for A and B (those values
depend on how the rule was created, i.e. what computation mode was chosen,
and especially the cardinality of the set A and B). By opposition, the notion
of contribution is defined to measure if individuals are more responsible for
the creation of the rule than the other ones. With the previous example, the
most contributive individuals are those who have 1 for the variables A and
B. So, the notions of typicality and contribution are different. In the same
way, the notion of typicality (resp. contribution) of a set of individuals (or of
a category of individuals) is defined in order to know if a considered set of
individuals is typical (resp. contributive) to a rule. In order to have formal
definitions of those notions, one should refer to the chapter of R. Gras and
P. Kuntz in this book.

7 An illustration with interval variables

and computation of typicality and contribution
This section is intended to give a simple and concrete example with the two
interval variables of Section 2. Figure 5 shows an implicative graph issued
from data of Figure 1. The two interval variables weight and height are au-
tomatically split into 4 intervals by CHIC as described in Section 2.

Fig. 5. An example of a implication graph

In the graph some interesting rules are visible. For example, we can see the
rules: weight1 ⇒ height12 and height34 ⇒ height2-4. Because the number of
individuals is small, and consequently not significant, and because the values of
this set have been arbitrarily generated, nothing else than: “light individuals
have generally a rather small height and that tall individuals are not light
ones” can be concluded. Nevertheless these rules show an implication between
the partitions of the two variables. Considering that these data can have a
52 Raphaël Couturier

sense for the expert, then we could have computed the typicality and the
contribution of the group of individuals. For example, concerning the rule
height34 ⇒ weight2-4, CHIC determines that the variable man contributes
the most to it (the error is 0.00638 for man, so close to 0; the error is equal to
1 for woman, so it does not contribute at all). In the opposite way the most
typical variable to rule weight1 ⇒ height12 is the variable woman (the error
is 0.00499 for woman, so this is a very good typicality; the error equals 1 for
man, so it does not contribute at all). Both those results are not surprising
analyzing the data.

8 Conclusion

CHIC implements almost all methods and techniques described in SIA. In

this chapter we have described the main features of CHIC. First the variables
that can be handled are described. These different kinds of variables enable
to model several particular cases frequently encountered. Some options of
CHIC are briefly given in order to allow users to know what they can do with
CHIC. Then, the three main representations are presented. The similarity tree
and the hierarchy tree respectively provide a non-oriented and an oriented
classification. The implication graph, which is the most interactive, allows
users to mine their data and highlights the important rules that may interest
the expert. For those three representations some functionalities are given.
Although the theory of SIA is far from being simple for a novice user, CHIC
allows to benefit from SIA results and distinguishes itself from other data
mining tools by its particular features.

References
1. R. Agrawal, T. Imielinski, and A. Swami. Mining association rules between
sets of items in large databases. In Proceedings of the 1993 ACM SIGMOD
International Conference on Management of Data, pages 207–216, 1993.
2. G. Bojadziev and M. Bojadziev. Fuzzy sets, fuzzy logic, applications. World
scientific, 1996.
3. R. Couturier. Un système de recommandation basé sur l’ASI. In Troisième
rencontre internationale de l’Analyse Statistique Implicative (ASI3), pages 157–
162, 2005.
4. R. Couturier, R. Gras, and F. Guillet. Reducing the number of variables us-
ing implicative analysis. In International Federation of Classification Societies,
IFCS 2004, pages 277–285. Springer Verlag: Classification, Clustering, and Data
Mining Applications, 2004.
5. E. Diday. La méthode des nuées dynamiques. Revue de statistique appliquée,
19(2):19–34, 1971.
6. G. Froissard. CHIC et les études docimologiques. In Troisième rencontre inter-
nationale de l’Analyse Statistique Implicative (ASI3), pages 187–197, 2005.
 CHIC: Cohesive Hierarchical Implicative Classification 53

7. R. Gras. Panorama du développement de l’A.S.I. à travers des situations fon-

datrices. In Actes de la 3ème Rencontre Internationale A.S.I., pages 9–33. Uni-
versité de Palerme, 2005.
8. R. Gras, S. Ag Almouloud, M. Bailleul, A. Lahrer, M. Polo, H. Ratsimba-
Rajohn, and A. Totohasina. L’implication Statistique. La Pensée Sauvage,
1996.
9. R. Gras, R. Couturier, J. Blanchard, H. Briand, P. Kuntz, and P. Peter.
Quelques critères pour une mesure de qualité de règles d’association. Un ex-
emple : l’implication statistique, chapter Mesures de qualité pour la fouille de
données, pages 3–32. RNTI-E-1, Cepaduès Editions, 2004.
10. R. Gras and P. Kuntz. Discovering R-rules with a directed hierarchy. Soft
Computing, A Fusion of Foundations, Methodologies and Applications, 1:46–58,
2005.
11. R. Gras, P. Kuntz, and J.C. Régnier. Significativité des niveaux d’une hiérarchie
orientée. Classification et fouille de données, RNTI-C-1, Cépaduès-Editions,
pages 39–50, 2004.
12. P. Lenca, P. Meyer, P. Vaillant, P. Picouet, and S. Lallich. Evaluation et analyse
multi-critères de qualité des règles d’association, chapter Mesures de qualité pour
la fouille de données, pages 219–246. RNTI-E-1, Cépaduès, 2004.
13. I. C. Lerman. Classification et analyse ordinale des données. Dunod, 1981.
14. P. Orus and P. Gregori. Des variables supplémentaires et des élèves “fictifs”,
dans la fouille didactique de données avec CHIC. In Troisième rencontre in-
ternationale de l’Analyse Statistique Implicative (ASI3), pages 279–291, 2005.
Assessing the interestingness of temporal rules
with Sequential Implication Intensity

Julien Blanchard, Fabrice Guillet, and Régis Gras

Knowledge & Decision (KOD) research team

LINA — FRE CNRS 2729
Polytechnic School of Nantes University, France
[email protected]

Summary. In this article, we study the assessment of the interestingness of sequen-

tial rules (generally temporal rules). This is a crucial problem in sequence analysis
since the frequent pattern mining algorithms are unsupervised and can produce huge
amounts of rules. While association rule interestingness has been widely studied in
the literature, there are few measures dedicated to sequential rules. Continuing with
our work on the adaptation of implication intensity to sequential rules, we propose
an original statistical measure for assessing sequential rule interestingness. More pre-
cisely, this measure named Sequential Implication Intensity (SII) evaluates the sta-
tistical significance of the rules in comparison with a probabilistic model. Numerical
simulations show that SII has unique features for a sequential rule interestingness
measure.

Key words: Temporal Data Mining, Event Sequences, Interestingness Measures

for Sequential Rules, Rule Significance.

1 Introduction
Frequent pattern discovery in sequences of events1 (generally temporal
sequences) is a major task in data mining. Research work in this domain
consists of two approaches:
• discovery of frequent episodes in a long sequence of events (approach
initiated by Mannila, Toivonen, and Verkamo [12, 13]),
• discovery of frequent sequential patterns in a set of sequences of events
(approach initiated by Agrawal and Srikant [1, 17]).
The similarity between episodes and sequential patterns is that they are
sequential structures, i.e., a structure defined with an order (partial or total).
Such a structure can be, for example:
1
Here we speak about sequences of qualitative variables. Such sequences are
generally not called time series.
J. Blanchard et al.: Assessing the interestingness of temporal rules with Sequential Implication
Intensity, Studies in Computational Intelligence (SCI) 127, 55–71 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
56 Blanchard et al.

breakfast then lunch then dinner

The structure is described by its frequency (or support) and generally by
constraints on the event position, like a maximal time window “less than 12
hours stand between breakfast and dinner ” [5, 10, 14, 17, 18].
The difference between episodes and sequential patterns lies in the mea-
sure of their frequency: frequency of episodes is an intra-sequence notion [5,
10, 14, 18–20], while frequency of sequential patterns is an inter-sequence no-
tion [1, 8, 16, 17, 21] (see [11] for a synthesis on the different ways of assessing
frequency). Thus, the frequent episode mining algorithms search for structures
which often recur inside a single sequence. On the other hand, the frequent
sequential pattern mining algorithms search for structures which recur in nu-
merous sequences (independently of the repetitions in each sequence). These
last algorithms are actually an extension to sequential data of the frequent
itemset mining algorithms, used among other things to generate association
rules [2, 9].
Just as the discovery of frequent itemsets leads to the generation of associ-
ation rules, the discovery of episodes/sequential patterns is often followed by a
sequential rule generation stage which enables predictions to be made within
the limits of a time window [5, 10, 14, 16–19, 21]. Such rules have been used to
predict, for example, stock market prices [5] or events in a telecommunication
network [14, 18]. A sequential rule can be for instance:
6h
breakfast −−−→ lunch
This rule means “if one observe breakfast then one will certainly observe lunch
less than 6 hours later”.
In this article, we study the assessment of the interestingness of sequential
rules. This is a crucial problem in sequence analysis since the frequent pattern
mining algorithms are unsupervised and can produce a huge number of rules.
While association rule interestingness has been widely studied in the literature
(see [3] for a survey), there are few measures dedicated to sequential rules. In
addition to frequency, one mainly finds an index of confidence (or precision)
that can be interpreted as an estimation of the conditional probability of
the conclusion given the condition [5, 10, 14, 16–19, 21]. A measure of recall is
sometimes used too; it can be interpreted as an estimation of the conditional
probability of the condition given the conclusion [18, 19]. In [5] and [10], the
authors have proposed an adaptation to sequential rules of the J-measure of
Smyth and Goodman, an index coming from mutual information2 . Finally, an
entropic measure is presented in [20] to quantify the information brought by
an episode in a sequence, but this approach only deals with episodes and not
with prediction rules.
These measures have several limits. First of all, the J-measure is not very
ω
intelligible since it gives the same value to a rule a −−−→ b and to its opposite
2
The J-measure is the part of the average mutual information relative to the truth
of the condition.
 Sequential Implication Intensity 57
ω
a −−−→ b, whereas these two rules make conflicting predictions. Confidence
and recall, vary linearly, which makes them rather sensitive to noise. Above
all, these measures increase with the size of the time window chosen. This
behavior is absolutely counter-intuitive since a rule with a too large time
window does not contribute to making good quality predictions. Indeed, the
larger the time window, the greater the probability of observing the conclusion
which follows the condition in data, and the less significant the rule. Another
major problem, which concerns confidence, recall, and J-measure, is that these
indexes are all frequency-based: the phenomena studied in data are considered
only in a relative way (by means of frequencies) and not in an absolute way
(by means of cardinalities). Thus, if a sequence is made longer by repeating it
x times one after the other, the indexes do not vary3 . Statistically, the rules
are all the more reliable since they are assessed on long sequences yet. In
the end, a good interestingness measure for sequential rules should therefore
decrease when the size of the time window is too large, and increase with
sequence enlargement. These essential properties have never been highlighted
in the literature.
Continuing with our work begun in [4], on the adaptation of implication
intensity to sequential rules [6, 7], we propose in this article an original statis-
tical measure for assessing sequential rule interestingness. More precisely, this
measure evaluates the statistical significance of the rules in comparison with a
probabilistic model. The next section is dedicated to the formalization of the
notions of sequential rule, example of a rule, and counter-example of a rule,
and to the presentation of the new measure, named Sequential Implication
Intensity (SII). In section 3, we study SII in several numerical simulations
and compare it to other measures.

2 Measuring the statistical significance of sequential

rules
2.1 Context

Our measure, SII, evaluates sequential rules extracted from one unique
sequence. This approach can be easily generalized to several sequences, for
example by computing an average or minimal SII on the set of sequences.
ω
Rules are of the form a −−−→ b, where a and b are episodes (these ones can
even be structured by intra-episode time constraints). However, in this article,
we restrict our study to sequential rules where the episodes a and b are two
single events.

3
We consider here that the size of the time window is negligible compared to the
size of the sequence, and we leave aside the possible side effects which could make
new patterns appear overlapping the end of a sequence and the beginning of the
following repeated sequence.
58 Blanchard et al.

The studied sequence is a continuous sequence of instantaneous events

(adaptation to discrete sequences is trivial). It is possible that two different
events occur at the same time. This amounts to using the same framework
as the one introduced by Mannila, Toivonen, and Verkamo [14]. To extract
the appropriate cardinalities from the sequence and compute SII, one only
needs to apply their episode mining algorithm named Winepi [13, 14] (or one
of its variants). In the following, we stand at the post-processing stage by
considering that Winepi has already been applied on the sequence, and we
directly work on the episode cardinalities that have been discovered. Here
again, our approach could be generalized to other kinds of sequences, for which
other episode mining algorithms have been proposed. For example, Höppner
has studied sequences with time-interval events that have a non-zero duration
and can overlap [10].

2.2 Notations

Fig. 1. A sequence S of events from E = {a, b, c} and its window F of size ω

beginning at TF .

Let E = {a, b, c, . . .} be a finite set of event types. An event is a couple

(e, t) where e ∈ E is the type of the event and t ∈ R+ is the time the event
occurred. It must be noted that the term event is often used to refer the event
type without reducing intelligibility.
An event sequence S observed between the instants Tstart and Tend is a
finite series of events
 
S = (e1 , t1 ), (e2 , t2 ), (e3 , t3 ), . . . , (en , tn )

such that:
 Sequential Implication Intensity 59

∀i ∈ {1..n}, (ei ∈ E ∧ ti ∈ [Tstart , Tend ])

∀i ∈ {1..n − 1}, ti ≤ ti+1
∀(i, j) ∈ {1..n}2 , ti = tj ⇒ ei 6= ej
The size of the sequence is L = Tend − Tstart .

A window on a sequence S is a subsequence of S. For instance, a window

F of size ω ≤ L beginning at the instant tF ∈ [Tstart , Tend − ω] contains all
the events (ei , ti ) from S such as tF ≤ ti ≤ tF + ω.

In the following, we consider a sequence S of events from E.

2.3 Sequential rules

We establish a formal framework for sequence analysis by defining the notions

of sequential rule, example of a rule, and counter-example of a rule. The ex-
amples and counter-examples of a sequential rule have never been defined in
the literature about sequences.

Fig. 2. Among the 3 windows of size ω beginning on events a, one can find 2
ω
examples and 1 counter-example of the rule a −−−→ b.

ω
Definition 1 A sequential rule is a triple (a, b, ω) noted a −−−→ b where
a and b are events of different types and ω is a strictly positive real number.
It means: “if an event a appears in the sequence then an event b certainly
appears within the next ω time units”.
ω
Definition 2 The examples of a sequential rule a −−−→ b are the events
a which are followed by at least one event b within the next ω time units.
Therefore the number of examples of the rule is the cardinality noted nab (ω):
 
0 0
 
nab (ω) = (a, t) ∈ S | ∃(b, t ) ∈ S, 0 ≤ t − t ≤ ω 

60 Blanchard et al.
ω
Definition 3 The counter-examples of a sequential rule a −−−→ b are the
events a which are not followed by any event b during the next ω time units.
Therefore the number of counter-examples of the rule is the cardinality noted
nab (ω):
 
0 0 0
 
nab (ω) = (a, t) ∈ S | ∀(b, t ) ∈ S, (t < t ∨ t > t + ω)


Contrary to association rules, nab and nab are not data constants but depend
on the parameter ω.

The originality of our approach is that it treats condition and conclusion in

very different ways: the events a are used as references for searching the events
b, i.e. only the windows which begin by an event a are taken into account.
On the contrary, in the literature about sequences, the algorithms like Winepi
move a window forward (with a fixed step) over the whole sequence [14]. This
method amounts to considering as examples of the sequential rule any window
that has an event a followed by b, even if it does not start by en event a. In
comparison, our approach is algorithmically less complex.

Let us note na the number of events a in the sequence. We have the usual
ω
equality na = nab + nab . A sequential rule a −−−→ b is completely described
by the quintuple (nab (ω), na , nb , ω, L). The examples of a sequential rule now
being defined, we can specify our measure for the frequency of the rules:

ω
Definition 4 The frequency of a sequential rule a −−−→ b is the proportion
of examples compared to the size of the sequence:

ω nab (ω)
f requency(a −−−→ b) =
L

With these notations, the confidence, recall, and J-measure are given by
the following formula:

ω nab (ω)
conf idence(a −−−→ b) =
na

ω nab (ω)
recall(a −−−→ b) =
nb
ω nab (ω) nab (ω)L nab (ω) nab (ω)L
J−measure(a −−−→ b) = log2 + log2
L na nb L na (L − nb )
 Sequential Implication Intensity 61

2.4 Random model

Following the implication intensity for association rules [6], the sequential
implication intensity SII measures the statistical significance of the rules
ω
a −−−→ b. To do so, it quantifies the unlikelihood of the smallness of the num-
ber of counter-examples nab (ω) with respect to the independence hypothesis
between the types of events a and b. Therefore, in a search for a random
model, we suppose that the types of events a and b are independent. Our
goal is to determine the distribution of the random variable Nab (number of
counter-examples of the rule) given the size L of the sequence, the numbers
na and nb of events of types a and b, and the size ω of the time window which
is used.
We suppose that the arrival process of the events of type b satisfies the
following hypotheses:
• the times between two successive occurrences of b are independent random
variables,
• the probability that a b appears during [t, t + ω] only depends on ω.
Moreover, two events of the same type cannot occur simultaneously in the
sequence S (see section 2.2). In these conditions, the arrival process of the
events of type b is a Poisson process of intensity λ = nLb . So, the number of b
appearing in a window of size ω follows Poisson’s Law with parameter ω.n L .
b

In particular, the probability that no event of type b appears during ω time

units is:
ω.nb ω
p = P(P oisson( ) = 0) = e− L nb
L
Therefore, wherever it appears in the sequence, an event a has the fixed prob-
ability p of being a counter-example, and 1 − p of being an example. Let us
repeat na times this random experiment to determine the theoretical number
of counter-examples Nab . If ω is negligible compared to L, then two randomly
chosen windows of size ω are not likely to overlap, and we can consider that
the na repetitions of the experiment are independent. In these conditions, the
random variable Nab is Binomial with parameters na and p:
ω
Nab = Binomial(na , e− L nb )

When permitted, this Binomial distribution can be approximated by another

Poisson distribution (even in the case of “weakly dependent” repetitions —
see [15]).

ω
Definition 5 The sequential implication intensity (SII ) of a rule a −−−→
b is defined by:
ω
SII(a −−−→ b) = P(Nab > nab (ω))
62 Blanchard et al.

Numerically, we have:
nab (ω)
ω ω ω
X
SII(a −−−→ b) = 1−P(Nab ≤ nab (ω)) = 1− Ckna (e− L nb )k (1−e− L nb )na −k
k=0

3 Properties and comparisons

Fig. 3. SII w.r.t. the number of counter-examples.

SII quantifies the unlikelihood of the smallness of the number of counter-

examples nab (ω) with respect to the independence hypothesis between the
ω
types of events a and b. In particular, if SII(a −−−→ b) is worth 1 or 0, then
it is unlikely that the types of event a and b are independent (deviation from
independence is significant and oriented in favor of the examples or of the
counter-examples). This new index can be seen as the complement to 1 of the
p-value of a hypothesis test. However, following the implication intensity, the
aim here is not testing a hypothesis but actually using it as a reference to
evaluate and sort the rules.
In the following, we study SII in several numerical simulations and com-
pare it to confidence, recall, and J-measure. These simulations point out the
intuitive properties of a good interestingness measure for sequential rules.

3.1 Counter-example increase

In this section, we study the measures when the number nab of counter-
ω
examples increases (with all other parameters constant). For a rule a −−−→ b,
 Sequential Implication Intensity 63

Fig. 4. SII, confidence, recall, and J-mesure w.r.t. the number of counter-examples.
na = 50, nb = 130, ω = 10, L = 1000

this can be seen as making the events a and b more distant in the sequence
while keeping the same numbers of a and b. This operation transforms events
a from examples to counter-examples.
Fig. 4 shows that SII clearly distinguishes between acceptable numbers
of counter-examples (assigned to values close to 1) and non-acceptable num-
bers of counter-examples (assigned to values close to 0) with respect to the
other parameters na , nb , ω, and L. On the contrary, confidence and recall
vary linearly, while J-measure provides very little discriminative power. Due
to its entropic nature, the J-measure could even increase when the number
of counter-examples increases, which is disturbing for a rule interestingness
measure.

3.2 Sequence enlargement

We call sequence enlargement the operation which makes the sequence longer
by adding new events (of new types) at the beginning or at the end. For a
ω
rule a −−−→ b, such an operation does not change the cardinalities nab (ω) and
nab (ω) since the layout of the events a and b remain the same. Only the size
L of the sequence increase.
Fig. 5 shows that SII increases with sequence enlargement. Indeed, for a
given number of counter-examples, a rule is more surprising in a long sequence
rather than in a short one since the a and b are less likely to be close in a
long sequence. On the contrary, measures like confidence and recall remain
unchanged since they do not take L into account (see Fig. 6). The J-measure
varies with L but only slightly. It can even decrease with L, which is counter-
intuitive.
64 Blanchard et al.

Fig. 5. SII with sequence enlargement.

na = 50, nb = 130, ω = 10

Fig. 6. SII, confidence, recall, and J-mesure with sequence enlargement.

na = 50, nb = 130, nab = 10, ω = 10

3.3 Sequence repetition

We call sequence repetition the operation which makes the sequence longer by
repeating it γ times one after the other (we leave aside the possible side effects
which could make new patterns appear by overlapping the end of a sequence
and the beginning of the following repeated sequence). With this operation,
the frequencies of the events a and b and the frequencies of the examples and
counter-examples remain unchanged.
Fig. 7 shows that the values of SII are more extreme (close to 0 or 1)
with sequence repetition. This is due to the statistical nature of the measure.
 Sequential Implication Intensity 65

Fig. 7. SII with sequence repetition.

na = 50 × γ, nb = 130 × γ, ω = 10, L = 1000 × γ

(a) nab = 12 × γ (b) nab = 16 × γ

Fig. 8. SII, confidence, recall, and J-mesure with sequence repetition.

na = 50 × γ, nb = 130 × γ, ω = 10, L = 1000 × γ
66 Blanchard et al.

Statistically, a rule is all the more significant when it is assessed on a long se-
quence with lots of events: the longer the sequence, the more one can trust the
imbalance between examples and counter-examples observed in the sequence,
and the more one can confirm the good or bad quality of the rule. On the
contrary, the frequency-based measures like confidence, recall, and J-measure
do not vary with sequence repetition (see Fig. 8).

3.4 Window enlargement

Fig. 9. A sequence where the events b are regularly spread.

Window enlargement consists of increasing the size ω of the time window.

As the function nab (ω) is unknown (nab is given by a data mining algorithm,
it depends on the data), we model it in the following way:

na nb L
nab (ω) = na − ω, if ω ≤
L nb
nab (ω) = 0 , otherwise.
This is a simple model, considering that the number of examples observed in
the sequence is proportional to ω: nab (ω) = naLnb ω. The formula is based on
the following postulates:
• According to definitions 2 and 3, nab must increase with ω and nab must
decrease with ω.
• If ω = 0 then there is no time window, and the data mining algorithm
cannot find any example4 . So we have nab = 0 and nab = na .
• Let us consider that the events b are regularly spread over the sequence
(Fig. 9). If ω ≥ nLb , then any event a can capture at least one event b within
the next ω time units. So we are sure that all the events a are examples,
i.e. na = nab and nab = 0.
In practice, since the events b are not regularly spread over the sequence, the
maximal gap between two consecutive events b can be greater than nLb . So the
4
We consider that two events a and b occurring at the same time do not make an
example.
 Sequential Implication Intensity 67

threshold ω ≥ nLb is not enough to be sure that na = nab . This is the reason
why we introduce a coefficient k into the function nab (ω):

na nb ω kL
nab (ω) = na − , if ω ≤
L k nb
nab (ω) = 0 , otherwise.
The coefficient k can be seen as a non-uniformity index for the events b in the
sequence. We have k = 1 only if the events b are regularly spread over the
sequence (Fig. 9).

Fig. 10. Model for nab (ω).

With this model for nab (ω), we can now study the interestingness measures
with regard to ω and k. Several interesting behaviors can be pointed out for
SII (see illustration in Fig. 11):
• There exists a range of values for ω which allows SII to be maximized.
This is intuitively satisfying5 . The higher the coefficient k, the smaller the
range of values.
• If ω is too large, then SII = 0. Indeed, the larger the time window, the
greater the probability of observing a given series of events in the sequence,
and the less significant the rule.
• As for the small values of ω (before the range of values which maximizes
SII):
– If k ≈ 1, then nab increases fast enough with ω to have SII increase
(Fig. 11 at the top).

5
When using a sequence mining algorithm to discover a specific phenomenon in
data, lots of time is spent to find the “right” value for the time window ω.
68 Blanchard et al.

– If k is larger, then nab does not increase fast enough with ω. SII
decreases until nab becomes more adequate (Fig. 11 at the bottom).
On the other hand, confidence (idem for recall) increases linearly with ω
(see Fig. 12 with a logarithmic scale). Above all, the three measures confi-
dence, recall, and J-measure do not tend to 0 when ω is large6 . Indeed, these
measures depend on ω only through nab , i.e. the parameter ω does not explic-
itly appear in the formulas of the measures. If ω is large enough to capture all
the examples, then nab = 0 is fixed and the three measures become constant
functions (with a good value since there is no counter-example). This behavior
is absolutely counter-intuitive. Only SII takes ω explicitly into account and
allows rules with too large time window to be discarded.

4 Conclusion

In this article, we have studied the assessment of the interestingness of sequen-

tial rules. First, we have formalized the notions of sequential rule, example of
a rule, and counter-example of a rule. We have then presented the Sequential
Implication Intensity (SII), an original statistical measure for assessing se-
quential rule interestingness. SII evaluates the statistical significance of the
rules in comparison with a probabilistic model. Numerical simulations show
that SII has interesting features. In particular, SII is the only measure that
takes sequence enlargement, sequence repetition, and window enlargement
into account in an appropriate way.
To continue this research work, we are developing a rule mining platform
for sequence analysis. Experimental studies of SII on real data (Yahoo Fi-
nance Stock Exchange data) will be available soon.

References
1. R. Agrawal and R. Srikant. Mining sequential patterns. In Proceedings of the
international conference on data engineering (ICDE), pages 3–14. IEEE Com-
puter Society, 1995.
2. R. Agrawal and R. Srikant. Fast algorithms for mining association rules. In
Jorge B. Bocca, Matthias Jarke, and Carlo Zaniolo, editors, Proceedings of the
twentieth international conference on very large data bases (VLDB 1994), pages
487–499. Morgan Kaufmann, 1994.
3. J. Blanchard. Un système de visualisation pour l’extraction, l’évaluation, et
l’exploration interactives des règles d’association. PhD thesis, Université de
Nantes, 2005.
4. J. Blanchard, F. Guillet, and H. Briand. L’intensité d’implication entropique
pour la recherche de règles de prédiction intéressantes dans les séquences de

6
This does not depend on any model chosen for nab (ω).
 Sequential Implication Intensity 69

Fig. 11. SII with window enlargement.

na = 50, nb = 100, L = 5000
70 Blanchard et al.

Fig. 12. Confidence with window enlargement.

na = 50, nb = 100, L = 5000

pannes d’ascenseurs. Extraction des Connaissances et Apprentissage, 1(4):77–

88, 2002. Actes des journées Extraction et Gestion des Connaissances (EGC)
2002.
5. G. Das, K.-I. Lin, H. Mannila, G. Renganathan, and P. Smyth. Rule discovery
from time series. In R. Agrawal, P. E. Stolorz, and G. Piatetsky-Shapiro, editors,
Proceedings of the fourth ACM SIGKDD international conference on knowledge
discovery and data mining, pages 16–22. AAAI Press, 1998.
6. R. Gras. L’implication statistique : nouvelle méthode exploratoire de données.
La Pensée Sauvage Editions, 1996.
7. R. Gras, P. Kuntz, R. Couturier, and F. Guillet. Une version entropique de
l’intensité d’implication pour les corpus volumineux. Extraction des Connais-
sances et Apprentissage, 1(1-2):69–80, 2001. Actes des journées Extraction et
Gestion des Connaissances (EGC) 2001.
8. J. Han, J. Pei, and X. Yan. Sequential pattern mining by pattern-growth:
Principles and extensions. In W. W. Chu and T. Y. Lin, editors, Recent Advances
in Data Mining and Granular Computing (Mathematical Aspects of Knowledge
Discovery), pages 183–220. Springer-Verlag, 2005.
9. J. Hipp, U. Güntzer, and G. Nakhaeizadeh. Algorithms for association rule
mining a general survey and comparison. SIGKDD Explorations, 2(1):58–64,
2000.
10. F. Höppner. Learning dependencies in multivariate time series. In Proceedings
of the ECAI’02 workshop on knowledge discovery in spatio-temporal data, pages
25–31, 2002.
11. M. Joshi, G. Karypis, and V. Kumar. A universal formulation of sequential
patterns. Technical report, University of Minnesota, 1999. TR 99-021.
 Sequential Implication Intensity 71

12. H. Mannila and H. Toivonen. Discovering generalized episodes using minimal

occurrences. In Proceedings of the second ACM SIGKDD international con-
ference on knowledge discovery and data mining, pages 146–151. AAAI Press,
1996.
13. H. Mannila, H. Toivonen, and A. I. Verkamo. Discovering frequent episodes in
sequences. In Proceedings of the first ACM SIGKDD international conference
on knowledge discovery and data mining, pages 210–215. AAAI Press, 1995.
14. H. Mannila, H. Toivonen, and A I. Verkamo. Discovery of frequent episodes in
event sequences. Data Mining and Knowledge Discovery, 1(3):259–289, 1997.
15. Sheldon M. Ross. Introduction to Probability Models. 2006. 9th edition.
16. M. Spiliopoulou. Managing interesting rules in sequence mining. In PKDD’99:
Proceedings of the third European conference on principles of data mining and
knowledge discovery, pages 554–560. Springer-Verlag, 1999.
17. R. Srikant and R. Agrawal. Mining sequential patterns: generalizations and
performance improvements. In EDBT’96: Proceedings of the fifth International
Conference on Extending Database Technology, pages 3–17. Springer-Verlag,
1996.
18. X. Sun, M. E. Orlowska, and X. Zhou. Finding event-oriented patterns in long
temporal sequences. In Kyu-Young Whang, Jongwoo Jeon, Kyuseok Shim, and
Jaideep Srivastava, editors, Proceedings of the seventh Pacific-Asia conference
on knowledge discovery and data mining (PAKDD2003), volume 2637 of Lecture
Notes in Computer Science, pages 15–26. Springer-Verlag, 2003.
19. G. M. Weiss. Predicting telecommunication equipment failures from sequences
of network alarms. In Handbook of knowledge discovery and data mining, pages
891–896. Oxford University Press, Inc., 2002.
20. J. Yang, W. Wang, and P. S. Yu. Stamp: On discovery of statistically impor-
tant pattern repeats in long sequential data. In Daniel Barbará and Chandrika
Kamath, editors, Proceedings of the third SIAM international conference on data
mining. SIAM, 2003.
21. M. J. Zaki. SPADE: an efficient algorithm for mining frequent sequences. Ma-
chine Learning, 42(1-2):31–60, 2001.
 Part II

Application to concept learning in education,

teaching, and didactics
Student’s Algebraic Knowledge Modelling:
Algebraic Context as Cause of Student’s Actions

Marie-Caroline Croset, Jana Trgalova, and Jean-François Nicaud

LIG Laboratory, MeTAH Team

46, Av. Felix Viallet
38031 Grenoble Cedex, France
{Marie-Caroline.Croset, Jana.Trgalova, Jean-Francois.Nicaud}@imag.fr

Summary. In this chapter, we describe a construction of a student model in the

field of algebra. For gathering the data, we have used the Aplusix learning envi-
ronment, which allows students to make freely calculation steps and records all the
students’ actions. One way to build and update the student model is to precisely
follow what the student is doing, by means of a detailed representation of cognitive
skills. We are interested in persistent and reproducible actions, i.e., the same action
done by a student in different algebraic contexts, rather than in a local student ac-
tion. For discovering patterns of student behaviours, we use a statistical implicative
analysis which makes possible seeking for stability of the actions and determining the
contexts where they appear. This theory allows us to build implicative connections
between algebraic contexts and actions.

Key words: Student model, Interactive Learning Environments, algrebraic trans-

formations, errors

1 Introduction

Teachers need information about students learning of reasoning processes in

order to be able to take appropriate didactical decisions: knowing where
the student has difficulties, what he/she masters, how his/her knowledge
evolves. . . They pay particular attention to errors: “errors are not only the
effect of ignorance, of uncertainty, of chance [. . . ], but the effect of a previ-
ous piece of knowledge which was interesting and successful, but which now
is revealed as false or simply not adapted” [1]. Usually established by hand
assignment, the collection and analysis of precise and individual information
about student’s knowledge is a slow and bothersome work, especially if there
is a large student body.
Interactive Learning Environments (ILEs) make it possible to overcome
these difficulties: automatic data collections are carried out and constitute
what is called a “student model”. Sison and Shimura define a student model
M.-C. Croset et al.: Student’s Algebraic Knowledge Modelling: Algebraic Context as Cause of
Student’s Actions, Studies in Computational Intelligence (SCI) 127, 75–98 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
76 M.-C. Croset et al.

as “an approximate, possibly partial, primarily qualitative representation of

student knowledge about a particular domain, or a particular topic or skill in
that domain, that can fully or partially account for specific aspects of student
behaviour” [2]. Student models not only inform teachers and researchers about
the students’ knowledge state, they also guide artificial tutors in the choice of
exercises to be presented to the students. They need first an accurate behav-
iour diagnosis and second a technique to analyze the resulting data, in order
to select which errors should be corrected, or mentioned. Indeed, a systematic
and repeated remediation cannot be considered: Sleeman has pointed out that
error-specific and automatic feedback is no more useful to the student than
generic remediation [3], unless to detect patterns in behaviours and cognitive
reasons for these behaviours. Brown and Van Lehn have developed a “Repair
theory” to provide procedures that will account for bugs [4]. The authors see
a bug appearance as an attempt at a repair when the student’s knowledge
leads to an impasse. They are concerned with systematic bugs that outline
regularities, which constitutes an interesting point of view.
Following the same idea, in order to avoid a systematic ineffective remedi-
ation we are looking for organized rather than isolated errors, in the particular
field of algebra. Our data are collected thanks to the Aplusix learning envi-
ronment (cf. section 3). We have chosen the statistical implicative analysis
(SIA) [5] and the CHIC software where this theory is implemented to outline
stabilities. The reasons for this choice are described in section 4. We assume
that the SIA allows us:
• to determine individually which tasks present patterns of stable errors
(section 7),
• to outline student stable behaviours in terms of implicative links between
algebraic contexts and actions (section 6),
• to model behaviour groups for a population of students (section 8).

2 Interactive Learning Environments for Algebra

Since 2003, our team has been engaged in research on student modelling in
the field of algebra; especially in the area of transformational (rule-based)
activities [6] such as expanding, collecting like terms, factoring, and solving
equations. Expression transformation is concerned with changing the form of
the expression or equation in order to maintain equivalence.
The choice of an ILE to gather data is crucial. Indeed, transformation steps
in most ILEs do not depend only on the student’s decisions: they also depend
on the degree of initiative left to the student by the environment. The existing
systems for algebra are quite different in terms of possible actions, ranging
from Computer Algebra Systems (CASs)1 , such as Maple or Derive, where
1
In a CAS, the transformations are made by the system and often the computer
itself selects the sub-expressions to which the rules are applied. In addition, and
 Student’s Algebraic Knowledge Modelling 77

the system solves transformational tasks, to ILEs such as Aplusix where the
student has to perform all the transformations with the given expressions.
To avoid combinatorial explosion some ILEs limit the student action to one
calculation step. The expected step is often very simple. Such is for example
the case of Cognitive Tutors [7]: each student action is required to be on an
interpretable path. At each step, the student action is compared to applica-
ble rules in the model and immediate feedback is conventionally provided. If
the student action matches one of the applicable rules, the tutor accepts the
action and applies the rule to update the internal representation of the prob-
lem state. If the student action does not match the action of any applicable
rule in the model, the action does not register and the tutor provides a brief
message in the hint window. In case of ambiguity about the interpretation of
the student action, the student is presented with a disambiguation menu to
identify the appropriate interpretation of the action. The feedback is immedi-
ate and immediate error correction is required. As a consequence, Cognitive
Tutors make it possible to follow the student very closely but, as Mc Arthur
points out, “because each incorrect rule is paired with a particular tutorial
action [. . . ], every student who takes a given incorrect step gets the same
message, regardless of how many times the same error has been made or how
many other errors have been made” [8].
Other authors ask the students to mark the rule they wish to apply. In
MathXpert [9] for example, the student selects a sub-expression, then chooses
a rule in a menu providing the rules that are applicable to this sub-expression,
see Fig. 1. The chosen rule is then automatically applied. The opportunity to
make mistakes in such an environment is strongly restricted. The T-algebra
environment [10] differs from MathXpert in the fact that it is the student
who has to write the result of the rule application (when the system is in ‘free
mode’). However, the rules menu presented by the system is also contextual:
it depends on the selected sub-expression.
A completely different approach is used in the Aplusix environment [11]:
this microworld leaves the students entirely free to produce the expressions
they wish, without specifying which rules they should apply. This environment
allows students to apply several rules in one single step, as they do in the
usual paper and pencil environment. Therefore, with such environments, we
are closer to the real mental processes of students. For this reason, we have
chosen Aplusix environment for our study of students’ actions.

contrary to the ILEs, the commands of a CAS are very powerful: simplify, fac-
torise, expand, solve, differentiate, integrate, and so on.
78 M.-C. Croset et al.

Fig. 1. Screenshot of MathXpert: rules menu is contextual to the selected expression

3 Aplusix Learning Environment

3.1 Presentation

Four different types of exercises, built by teachers or researchers, can be pro-

posed to students: calculate numerical expressions, expand and reduce poly-
nomial expressions, factor polynomial expressions, and solve equations, in-
equalities or systems of linear equations. In each situation, an expression and
an instruction are given to the students. To solve the exercise, the students
can duplicate the expression and modify it, see Fig. 2. The transformation
of an expression into another one is called a student’s step. In the test mode
no feedback is given to students, while in the training mode epistemic feed-
back is generated in terms of indicators giving the state of expressions and
the correctness of the student’s calculations. Aplusix records in text files the
following information about the student’s actions: time, keyboard or mouse
actions, and expression obtained. The files can be viewed by the student, the
teacher or the researcher thanks to a ‘replay system’ included in the software.

Fig. 2. Screenshots of Aplusix. On the left, the student is in test mode, he/she has
done several transformations in each step. On the right, the student is in training
mode, with feedback about correctness
 Student’s Algebraic Knowledge Modelling 79

Aplusix permits the occurrence of complex errors and actions in one stu-
dent’s step, as shown in Fig. 2. As a result, the understanding of the student
reasoning is complicated and the difficulty of providing a diagnosis of mistakes
increases. It is then necessary to subdivide a student’s step into elementary
steps. This is done by an automatic process: the rules diagnosis provided by
Anaïs, presented in the following section.

3.2 Rules Diagnosis

The files used for an automatic analysis consist only of the calculation steps
validated by the student (i.e., the student’s steps): corrections, hesitations and
time are not taken into account. Software, called Anaïs, has been developed to
analyse students’ productions. It is based on rules established from a didactical
analysis and gathered in a library. The analysis consists of searching for the
best sequence of the rules (correct or incorrect) that can explain a given
student’s step. The process of the analysis is as follows: from the expression
that is the source of the student’s step, Anaïs develops a tree by applying all
the rules applicable to this expression.
• The application of a rule produces a new node. Anaïs thus gradually builds
a research tree, at each level choosing the node to be developed by using
a heuristic that takes into account the goal (the expression resulting from
the student’s step).
• When the process is successful, the goal may be reached by several paths,
each of them being a diagnosis. The selection of the best diagnosis is
based on a cost of paths expressed in terms of the number and kind of
rules applied.
The Anaïs software provides a diagnosis in the form of a sequence of in-
termediate elementary stages (rewriting rules) to explain the steps produced
by the student, as shown in Tab. 1.
We call elementary step an automatic intermediate stage provided by
Anaïs. An elementary step has an initial expression and a final (intermediate)
expression.
A single rule explains each elementary step. A step can belong to one of
four different tasks, whatever the type of the exercise: expansion, factoring,
collecting like terms, and movement, cf. Tab. 1. An exercise of the equation-
solving type may involve factoring, collecting like terms, and movement steps.

Remark 1. One can question the interest of leaving such a large freedom to
the student’s answers and thus having to reconstruct the intermediate steps.
Indeed, the ILEs presented in section 2 do not need to do this complex work.
However, when a student solves an algebra exercise, he/she does not always
see an expression transformation as the application of a rule, with an initial
and a final expression. Requiring that students cite the rules applied may be
80 M.-C. Croset et al.

(Automatic) Initial Final Associated Associated

Student’s step
elementary steps expression expression automatic rules Task
ax + bx 7→ (a − b)x
7x − 2x + 4 = 0 7x − 2x 7→ 9x 7x − 2x 9x Reduction
(incorrect)
9x + 4 = 0 a + b = 0 7→ a = −b
7→ 9x = −4 9x + 4 = 0 9x = −4 Movement
7→ 9x = −4 (correct)

Table 1. Example of a student’s step decomposition into automatic elementary

steps

of didactical interest, but it may lead us away from the student’s real way of
thinking. The freedom the students have while working with Aplusix puts us
as close as possible to their real mental processes.

4 The Choice of Statistical Implicative Analysis

One of the difficulties in catching stable errors is to define what stability is:
what is a good threshold to decide when a rule can be considered as having
been regularly used? The most common and spontaneous technique for catch-
ing stability is to count the number of opportunities for a rule application and
divide it by the number of effective student’s applications [12,13]. Authors do
not define very precisely what is called opportunity for a rule application.
They recognize themselves that “different mal-rules have widely different ‘op-
portunities’ of occurring, and in some cases the number of opportunities is
impossible to quantify without looking closely at individual students proto-
cols” [13].

4.1 Difficulties in Defining what Stable Behaviours Are

Let us take an example to show the difficulty of stability definition. Two

students, A and B, are asked to collect like terms in the five expressions given
in Tab. 2.

3x − 5 + 3x E1
3x + 3x − 5 E2
−5 + 3x + 3x E3
7x − 8 + 3 + 7x E4
5x × 2x E5

Table 2. Expressions proposed to student A and B

Let us call R, the rule: ax + bx 7→ (a − b)x.

 Student’s Algebraic Knowledge Modelling 81

The four expressions E1, E2, E3, and E4 present an opportunity to apply
R. Clearly, E5 does not present an opportunity for application of R because
the initial expression is a product instead of a sum. Therefore, it seems that
there are four opportunities for application of R. Let us see what the answers
of students for the four expressions are:
Let us suppose that student A’s answer is −5 for each of the first four
expressions (and, for example, 10x2 , for E5). This means that student A has
used R for these four expressions: the frequency of application of R associated
to student A is then 4/4. It seems that the behaviour of student A is very
stable with respect to this rule.
Let us suppose now that student B’s answers are respectively: −5, 6x − 5,
−5 + 6x and −5 for the first four expressions (and, for example, 10x2 , for
E5). This would mean that the student B has used R only for E1 and E4.
The frequency of application of R for the student B would then be 2/4. Does
it mean that the application of R for the student B is unstable? We do not
believe that.
Balacheff explains that two different behaviours (using two different rules
here) “can appear as conflicting but this inconsistency can be explained either
by the time evolution or by the situation/context” [14]. The algebraic context
of the four expressions is not the same: in E1 and E4, the minus sign is between
the two monomials. Moreover, the minus sign is placed side by side with 3x,
respectively 7x. In the expressions E2 and E3, there is also a minus sign, but it
is not between the two equal monomials. The didactical variable, which is the
position of the minus sign in the expressions, does not take the same values for
each expression, and it influences the behaviour of student B. We can say that
the behaviour of student B is stable, with respect to the minus sign position:
for him/her, the expressions E2 and E3 do not present an opportunity to
apply R. Each student has his own conception of the opportunity for applying
the rule R. For this reason, it does not seem possible to objectively count the
number of opportunities for a rule application. The application opportunity
is a subjective notion: a “transfer from one situation to another one is not
an obvious process, even if in the eyes of an observer these situations are
isomorphic” [15]. We see that stability definition depends on what is called
opportunity.

4.2 Algebraic Context as Source of Behaviour

In our work, we decided not to have a priori ideas of what constitute oppor-
tunities for a given rule application. We suppose that the sources of errors in
incorrect transformations are principally in the characteristics of the expres-
sion, such as the degree of the initial expression, the nature of its coefficients,
the presence of a minus sign, and so on. We need to describe precisely the
algebraic characteristics of the initial expression to which a rule is applied
and we are looking for the algebraic context that can ‘better’ explain the rule
utilisation for each student.
82 M.-C. Croset et al.

4.3 Statistical Implicative Analysis

Since we are looking for causes of students’ errors and not just correlations, the
choice of SIA seems worthwhile. Indeed, the SIA approach makes it possible
to find implicative links between attributes, in our case between algebraic
characteristics and rules. In addition, this technique takes into account the
number of times that a context appears relative to the other contexts. For
example, if a student S1 uses a rule R1 ten times in a context C0 , and he/she
uses a rule R2 five times in the same context C0 , then the quasi-implication
C0 → R1 can be evaluated.2 Let us consider another student, S2 , who uses
the rule R1 100 times in the context C0 and the rule R2 50 times in the same
context. The frequency of R1 application by both students is the same (2/3),
but the SIA approach makes a distinction between 10/15 and 100/150.

5 The Conditions of the Expected Quasi-Implications

5.1 Algebraic Context Variables

Let us consider a rule set, {Rk }1≤k≤p .3 To this set, we associate algebraic con-
text variables, {Vi }1≤i≤n that are the main characteristics of initial expressions
on which each rule Rk can be applied. Each algebraic context variable, Vi , (just
called variable in what follows) can be assigned values noted {V ij }1≤j≤mi .
We call contexts the vectors (V 1j1 , V 2j2 , . . . , V njn ), associated to an ex-
i=n
Y
pression. The number of such vectors is mi . We denote a context by Cl ,
i=1
i=n
Y
where l ∈ {1, . . . , mi }.
i=1

Remark 2. Since we will use the CHIC software, we prefer to have variables
with binary values. Therefore, we consider the values {V ij }1≤j≤mi as new
variables, called binary context variables, which take 0 or 1 as values. When
necessary, the distinction between binary context variables and algebraic con-
text variables will be made.
We illustrate this with an example. Let us consider the expression 3x − 5.
The operator and the presence of a minus sign are two algebraic context vari-
ables of the expression. The operator can be one of the five binary context
variables which are times, plus, minus (e.g. the expression −(2x + 7)), bracket
and exponent. These binary context variables can be assigned binary values:
2
Quasi-implication is called a “rule” by the authors of the SIA approach. In order
not to confuse it with what we call algebraic rules, we will call it implication or
quasi-implication, noted →, while the transformation of an algebraic expression
is noted 7→.
3
In general, rule set is associated to a task or a part of task.
 Student’s Algebraic Knowledge Modelling 83

the expression has or does not have the particular operator. The second vari-
able, presence of a minus sign, has one binary variable: itself. A context extract
of the expression 3x − 5 is then (Plus, Presence of minus sign).
Obviously, variables depend on tasks: variables for factoring or for move-
ment are not the same. We assume that the variables have impact on the use
of a rule by a student. This means that the rule used depends on the value of
a variable.
We will consider a behaviour as stable if the student uses the same rule
each time (or almost each time) that he/she is in the same algebraic context.
Description and choice of variables are determining factors for catching sta-
bility and this is a difficult task. The values depend on didactical decisions,
as we will show in the next subsection.

5.2 Creation of the Algebraic Context Variables List

The choice of the main characteristics of an expression is based on a two-steps

procedure:
• a praxeological analysis [16] of textbooks in the field of transformational
activities,
• a didactical analysis, together with the construction of the rules library.
Chevallard, in his Anthropological Theory of Didactics (ATD), describes
mathematical knowledge both as a means and a product of activity, which
form the “praxeological organization”, a union of practice and discourse about
practice. The basic elements of the anthropological model (types of prob-
lems, techniques, technologies and theories) make it possible to analyze math-
ematical textbooks and organize problems by types. What makes a difference
between two types of problems constitutes our first list of variables. For exam-
ple, two types of problems related to the task of expansion are expand a(b + c)
and expand (a + b)(c + d). One of the variables can thus be the number of the
terms in the sum. For example, in the first case, this variable will be assigned
the value (1, 2) while in the second case, the value (2, 2).
This work is not sufficient. Indeed, if we limit ourselves to textbook analy-
sis, we remain within the institutional contract. One thing that can cause
errors is a rupture of this contract. Therefore, we need to complete the first
list with what we think may be missing in the textbooks. For example, in the
task of expansion very few textbooks propose problems with three factors. We
have then considered the number of factors as a new variable associated to
the task of expansion.
In addition, we have to describe what the values of the variables are and
which are relevant for a change of the student’s behaviour. For example, the
degree of an expression is obviously a variable. But what values are pertinent
for this variable? When interviewed, some students pointed out that they
knew how to deal with expressions of degree 2 but not with expressions of
degree 3 and more. This leads us to suppose that the fact that an expression
84 M.-C. Croset et al.

is of degree 3 or 4 or 5 makes no difference for a student. The values of the

expression degree will then be 0, 1, 2 and greater than or equal to 3.
Not all these variables are as yet what Brousseau calls didactical vari-
ables [17]: we are not sure that they provoke a change of strategy or a change
of rules use. However, we think that they can have an impact on strategies for
some students. The SIA analysis will say which of the variables are didactical
and which are not. Indeed, the SIA analysis will give results in terms of im-
plications between variables and actions. The variables that emerge in these
implications will be considered to be didactical variables.

5.3 Presentation of the Files Analysed

Experimentation with Aplusix has been carried out for different purposes
by teachers and researchers [18]. However, experimentation presented in this
chapter was all conducted in the test mode (i.e., without information about
correctness of students’ steps, see section 3). Log files were gathered in a
database and analysed by Anaïs to provide a sequence of elementary steps for
each student’s step.
We can query (with the Structured Query Language — SQL) the database
and get result sets. These sets are tables which can be recorded in the CSV
format and directly used by the CHIC software. We will call these tables CHIC
tables.
The lines of the CHIC tables used in section 6 and 7 consist of the elemen-
tary steps from the automatic diagnosis for individual students. At least one
CHIC table is associated to one student. The columns, called attributes, con-
sist of the binary context variables and the actions. The actions can be either
the rules diagnosed by Anaïs (section 6) or a collection of rules (section 7),
according to what we want to model: a precise task or a set of tasks. The
values assigned to the rules are binary: either the rule is used or not. If an
elementary step in line i is explained by the rule which is in column j, there is
a 1 in the cell (i, j). The binary context variables are, as their name indicates,
binary. There is 1 in the cell (i, k) each time the initial expression of the step
i can be described by the (binary) variables of column k. An example of line
is shown in Tab. 3.
The files used in section 8 are not exactly the same and will be explained
in time.
Remark 3. On a line, it is not possible to have 1 in the columns of two binary
variables that depend on the same algebraic context variable while there can
be as many 1’s as there are context variables.

5.4 Implications

Four kinds of implications between attributes, binary context variables, V ij

and rules, Rk , are envisaged. Let us explain what they mean:
 Student’s Algebraic Knowledge Modelling 85

Elementary R1 : R2 : Decimal Integer Plus

Steps ax + bx 7→ (a − b)x ax + bx 7→ (a + b)x coefficients coefficients operator
7x − 2x 7→ 9x 1 0 0 1 1

Table 3. Extract of CHIC table analysed by the software CHIC

• V ij → Rk . This implication means that when expression variables Vi take

the value V ij , the rule Rk is almost always used by the student. This is
the most important implication for us.
• Rk → V ij . This implication means that when the rule Rk is used,V ij is
often the context in question. The contrapositive is easier to understand:
when the context is not V ij , Rk is not used.
• V ij → V rs , with r 6= i4 . This implication means that the algebraic vari-
able V ij mathematically implies the variable V rs . This occurs when an
expression in the data can be described both by V ij and V rs . For exam-
ple, in grade 8, students do not know how to solve equations of degree 2
in the canonical form, i.e. ax2 + bx + c = 0. However, they can solve such
equations if the left member is a product of two linear factors and the
right member is 0. The variables ‘degree 2’ and ‘member 0’ would be then
strongly correlated.
• Rk → Rs . This implication does not appear in our work. Each line is a
single elementary step: it means that, in the line, there is one and only
one rule in the whole set of rules that can be used.
An option in the CHIC software makes it possible to consider only the first
two implications: we select the rules, as the ‘principal vertex’ of the graph (see
Couturier chapter).

6 An Accurate Student’s Model: the Case of Factoring

Experiments were conducted in the field of factoring. They were built in order
to collect data on this domain. Of course, only the elementary steps of the
factoring task are collected in the CHIC table for analysis: steps concerning
collecting like terms or expanding are not taken into account. Here, the actions
are the rules associated to the elementary steps from the automatic diagnoses.
We have modelled students individually: the CHIC table represents here the
data collected about one student.

4
The index r cannot be equal to i: a value of a variable cannot imply another value
of the same variable.
86 M.-C. Croset et al.

6.1 The Attributes

The Actions.

In the rules library, there are 20 rules concerning factoring. We present here
the results for one student. Five rules have been diagnosed for this student:
• Correct: the factoring is correct.
• ErMinus: the factoring is erroneous and the mistake is about a minus sign.
For example, (5x + 1)x − (1 + 5x)y 7→ (5x + 1)(x + y).
• ErNothing: when the cofactor of the common factor is 1, some students
think that there remains “nothing” when the common factor is withdrawn.
For example, (x + 3)(x + 2) + (x + 3) 7→ (x + 3)(x + 2). This transformation
can be explained by the loss of a term but interviews with students show
that the concept of “nothing” was behind this transformation.
• ErOther: other kind of factoring errors.
• NoInf, meaning NoInformation: the student has not answered this task.
Note that it is not a step but an expression: there is no final expression
because the student has stopped solving the task.

The Context Variables.

There are 36 binary context variables associated to the factoring task. The
six context variables are the nature of the common factor, its visibility, its
degree, its position, the nature of its cofactors and the presence and position
of a minus sign. Each variable is decomposed into binary variables. In what
follows, we explain only those that appear in the implicative graph. Each of
them is illustrated by an example.
• The nature of the common factor can be numeric (e.g., 6x + 3), monomial
(e.g., 3x + 15x2 ), a sum of two terms (e.g., (5x + 1)x − (1 + 5x)y), a sum
of three terms (e.g., x(x + 2 + y) − (x + y + 2)(1 + 5x)), or a product (e.g.,
x(x − 4) + (x2 − 4x)(x + 1)).
• The visibility of the common factor depends on its nature. Let us take the
example of a sum of two terms as a common factor. Its visibility can be
obvious (e.g., (x+3)(x+2)+(x+3)), commuted (e.g., (5x+1)x−(1+5x)y),
opposite (e.g., (x + 3)(x + 2) + (−x − 3)), commuted-opposite (e.g., (x + 3)
(x + 2) + (−3 − x)), disconnected (e.g., x + (x + 2) × x + 2), multiple (e.g.,
−6 − 3x + (1 + x)(−2 − x)) or bi-multiple (e.g., −6 − 3x + (1 + x)(−4 − 2x)).
• The nature of the cofactors can be numeric, unit, monomial, sum, product
or identical. For example, in the expression (x + 3)(x + 2) + (x + 3), the
nature of the cofactors is respectively sum (x + 2) and unit (1); while in
the expression x(x − 4)(x + 1) + (x − 4)2 , the cofactors of the common
factor (x − 4) are on the one hand the product x(x + 1) and on the other
hand, the sum (x − 4), which is identical to the common factor.
 Student’s Algebraic Knowledge Modelling 87

6.2 The Experimentation

Contrary to the cases presented in sections 7 and 8, this experimentation car-

ried out with grade 8 students was organized especially to collect data about
factoring. Therefore, the exercises were built in order to cover a maximum
of types of the above mentioned variables. 10752 exercises could have been
created by taking the direct product of values of the variables. Of course, it
is not realistic to hope to find students who will work on so many exercises.
Moreover, some combinations are not interesting from a didactical point of
view: for example, it is not interesting to have two units for the two cofac-
tors, as in the expression (x + 5) + (x + 5). Moreover, if the common factor
is a product, we have decided not to accept a product as cofactor in order
not to generate excessively complicated expressions. For example, we have
not considered expressions like x(x + 1) + x(x + 1)(x + 2)(x + 4), where the
common factor is a product x(x + 1) and one of the cofactors is also a product
(x + 2)(x + 4). For these reasons, only 41 expressions were kept (see Tab. 4
for an extract). The student whose results are presented here worked for 70
minutes on these expressions without any help or feedback. She tried to solve
the whole range of the 41 exercises.

Examples Common factor Cofactors

FNumericObvious
FMonoObvious

Fsum2Commut

Expected value
Fsum2Obvious

Distribution
Distribution
Distribution
Distribution
Distribution

Distribution
Distribution
FSum2Opp
FSum2

Expressions
2y 2 + 2 1 0 0 0 0 0 0 0 0 1 0 1 0 0
−7x − 3x2 0 0 0 0 0 0 0 0 0 0 1 0 1 0
3x2 + yx 0 1 0 0 0 0 0 0 0 0 0 1 1 0
3(x + 8)2 + x2 + 8x 0 0 1 0 0 0 0 1 0 0 0 1 1 0
(x − 3)(1 − 4x) − 5(3 − x) 0 0 1 0 0 0 1 0 0 0 1 0 0 1
−6 − 3x + (2 − x)(−4 − 2x) 0 0 1 0 0 0 0 0 1 0 1 0 0 0
(x − 4) + (x − 4)x 0 0 1 1 0 0 0 0 0 1 0 1 0 0
−(x + 3 + y) + x(y + x + 3) 0 0 0 0 0 0 0 0 0 1 0 1 0 0

Table 4. Extract of the Experimentation Exercises. The Case of Factoring

6.3 The Results

The implicative graph was then built from a file with 41 lines and 36 + 5
attributes. A first work with one premise is presented; a second one follows
with two premises.
88 M.-C. Croset et al.

One premise.

Three sets of implications emerge from the analysis and are presented in Fig. 3:
• The student did not transform the expression, NoInf, when the common
factor or one of the cofactors is a product (FProduct and CofProduct). For
example, she did not treat exercises like x(x + 2)(x + 1) + 3(1 + x)(2 + x),
where the common factor is a product. She also did not answer when the
common factor is a disconnected sum (FSum2TermDisc), or an opposite
one (FSum2TermOpp). For example, the expression (x + (x + 2) × x + 2)
and −5x − 8 + (5x + 8)(x + 2) contributed respectively to the implication
(N oInf → F Sum2T ermDisc) and (N oInf → F Sum2T ermOpp). These
results are interesting. Even when the student does not answer, it means
something at a cognitive level: considering a product as a common factor
seems too difficult for this student.
• The student performed a correct factoring when the common factor was
numeric (Fnum), or monomial (Fmono). For example, the student correctly
factorized the expression −3y − 12 or 3x + 9x(x + 2).
• No interesting information about the ErNothing rule use emerges at this
level. A necessary condition for the use of ErNothing is the presence of
a unit cofactor in the expression (CofUnit). Indeed, this rule cannot be
applied if the cofactor is not a unit. It is this information that appears in
the implicative graph. We will see that with two premises we have more
information about the use of this rule.
We also added a new action, called Transformation that is the oppo-
site of the action NoInf. It means that the student transforms an expres-
sion either correctly or not. The following implications emerge at threshold
89 (see Couturier chapter): (F sum2T ermObvious → T ransf ormation) and
(F sum2T ermCommut → T ransf ormation). When the common factor is a
sum, either obvious or commuted, the student tries to transform the expres-
sion, the transformation being correct or not.

Fproduct CofProduct FSum2TermOpp FSum2TermDisc FNum FNum1Fact FMono1Fact ErNothing

NoInf Correct Correct

Fig. 3. Implicative graph with one premise. Threshold at 86. Case of factoring
 Student’s Algebraic Knowledge Modelling 89

Two premises.

The implications are more precise. One of the implications with two premises
is interesting to describe. The student uses the rule ErNothing especially
in the cases where the common factor is an obvious sum of two terms,
Fsum2TermObvious, and of course, one of the cofactors is a unit Cofunit,
cf. Fig. 4. The transformation (x + 3)(x + 2) + (x + 3) 7→ (x + 3)(x + 2) con-
tributes to this implication. When looking at the data, we see that indeed,
the student did not use this rule when she was confronted with an expression
like 5x + 5, even if the last cofactor is also a unit.

CofUnitFSum2TermObvious

ErNothing

CofUnit

Fig. 4. Part of the implicative graph with two premises. Threshold at 90. Case of
factoring

The study of the data of this student shows that she has stable behaviour
in the field of factoring. She has correct and incorrect stable actions, well
fixed in her behaviour. The SIA approach made it possible to detect stable
behaviours according to contexts. It answered one of our questions.

7 Detection of a Task that is the Source of a Student’s

Errors
To model a student’s behaviours, one needs to obtain enough data from this
student. This is one of our difficulties. This difficulty was overcome in the
previous analysis (section 6) by building special experiments about one precise
task. But teachers rarely have the possibility of leaving their students at work
on the same task for a long time. In addition, students learn while they work:
their knowledge evolves. We cannot model students after too long a time of
activities if we want to determine a state of their knowledge. Which task can
we, then, model with a set of exercises? Do we have to model each task? And
if so, do we have enough data? The SIA allows to answer these questions.
It makes it possible to detect which task causes the most errors in a set of
tasks, whatever the exercises are. For that, the process is quite similar to the
90 M.-C. Croset et al.

previous one: the CHIC table concerns only one student but it includes many
tasks done by her/him. The lines of the CHIC table are again elementary
steps, the columns are binary variables and actions. However, in contrast to
section 6, the actions are not the rules but sets of rules.

7.1 The Attributes

The Actions.

For this analysis, actions take only two values: either the elementary step is
correct (called Correct) or it is not (called Error). No distinction between the
kinds of errors is made.

The Context Variables.

Given an expression, there are 15 binary variables associated to these two

actions, described by six variables: the degree of the expression, the operator,
the nature of coefficients, the presence of minus sign, the presence of exponent
applied to a number and the task associated to the elementary step. The six
variables are explained in what follows:
• The degree of the expression, InDeg, can be assigned infinitely many values.
We assume that there are only four values that have different impact on
student’s behaviors: the degree 0, 1, 2 and greater than or equal to 3.
• The operator, InOp, can be: plus, times, exponent, parentheses, minus.
The expression x − 2 has a plus operator while the expression −(x + 3)
has a minus as operator.
• The coefficients can be integer, WIntCoef, fractional, decimal or irrational.
• The fact that there is a minus sign in the expression, WMinus, as in the
expression x − 2.
• The fact that there is exponent that is not applied to variables, as in the
expression x + 32 .
• The tasks associated to an elementary step are those described in section 3:
expansion, factorisation, collecting like terms, Collect, and movement. The
task information concerns the step much more than it does the source
expression. However, we prefer to consider it as a context variable rather
than an action.

7.2 Some Results

The treated CHIC table can contain as many elementary steps as the student
has done. In the following example, 99 steps have been diagnosed. The table
is then a 99 × 15 matrix. The student was chosen from an experiment that
especially concerned movement tasks. The results of the implicative graph are
presented in Fig. 5. This analysis shows that this student (grade 9) seems to
 Student’s Algebraic Knowledge Modelling 91

master the movement task: the implication M ovement → Correct appears.

However, he does not master the task of collecting like terms, especially when
the expression contains a minus sign and integer coefficients. This information
appears at threshold 93.

Movement InDeg1 Movement

WIntCoef Movement WIntCoef InDeg1 Movement

Correct

WIntCoef Collect InOpMinus WIntCoef InDeg1 InOpMinus

Collect InOpMinus Collect InDeg1 InOpMinus

Error

InOpMinus

Fig. 5. Implicative graph with three premises, threshold 87. Example of a result
for detecting a task that is the source of the student’s errors. In this case, the task
of collecting like terms, especially in the presence of a minus sign, seems to cause
regular difficulties for this student

This treatment answers to our previous questions: the SIA makes it possi-
ble to detect tasks that do not present difficulties to the student and those that
do. The student in question has difficulties with the specific task of collecting
like terms. It is this task that can be interesting to model, as it was done in
section 6: selection of the elementary steps concerning this task among the
whole 99 steps, association of the context variables in this task and analysis
of the resulting implicative graph. The last part can sometimes be complex.
In the next section, we will try to overcome this problem.

8 Seeking for Main Behaviours

The analysis of the factoring implicative graph presented in section 6 (Fig. 3),
is quite simple for those who have taken time to understand each attribute.
On the other hand, it is not possible to give this kind of result to a teacher
who would have as many graphs as students in his/her classroom. In addition,
some results can be much more complex. An expert cannot be present all the
time to analyze the implicative graph associated to each student. For this
reason, we try to detect the main possible behaviour groups for a given task
in order, in the near future, to automatically associate a student’s work to
one of these groups that would already be analysed.

8.1 Behaviour Groups

Previous CHIC tables have elementary steps as lines, actions and context
variables as columns. For detecting main behaviours, we proceed in another
way.
92 M.-C. Croset et al.

This time, we consider a population of n students. The attributes are

the ordered pair (Cl , Rk ), where Cl is a context (see section 5.1), and Rk a
rule. For each student S, a first new CHIC table, TS , is constructed in which
the lines are again elementary steps, but the columns are the ordered pairs
(Cl , Rk )5 . Let (Cl , Rk ) be the r-column. There is 1 in the cell (q, r) if the
context Cl describes the elementary step of the line q and the rule Rk is the
rule automatically associated to the step.
Let us consider a student S. The frequency, FS (Cl , Rk ), of the occurrence
of the ordered pair (Cl , Rk ) for the student S, is defined by the number of
times where Rk was used by the student S in the context Cl divided by the
number of times where the student S was confronted with the context Cl .
Remark 4. If the student S has never met the context Cl , the frequency
FS (Cl , Rk ) is equal to the mean of the frequencies of (Cl , Rk ) of the whole set
of students.
A second new CHIC table is created: The columns are again the ordered
pairs (Cl , Rk ) while each line concerns a particular student. The value of the
cell (q, r) is the frequency of the ordered pairs (Cl , Rk ) for the student who
is in the line q. If there are n students in the considered population, then
there are n lines in the CHIC table. The number of columns is the number of
i=n
Y
possible ordered pairs, namely p mi .
i=1
We are looking for groups of ordered pairs (Cl , Rk ). We call a behaviour
group a set of (Cl , Rk ) that is representative of a part of the population. For
that, we use a similarity tree viewable by the CHIC software: the attributes are
grouped according to the similarity of their use by the considered population.
For example, a behaviour group consisting of (C1 , R2 ) and (C3 , R1 ) means
that a group of students have used the rules R2 and R1 in the context C1 , or
respectively C3 .

8.2 The Case of Collecting Like Terms

We are concerned by a part of the task of collecting like terms: the sum
of two terms of the same degree, when the degree obtained is correct. In
another words, we consider only elementary steps like axm + bxm 7→ cxm ,
where m can be null, a and b can be positive or negative. We do not consider
transformations like axm +bxm 7→ cxp , where p is not equal to m. In addition,
we consider only four possibilities for the coefficient c: c is obtained as a sum
depending on a and b described below. We set aside the case where c is equal
to a or b, or ab, and so on. We look for behaviour groups for this sub-task.

5
If we take the notation used in the context definition (section 5.1), there are
i=n
Y
p mi couples.
i=1
 Student’s Algebraic Knowledge Modelling 93

The Rules.
The actions correspond to the four rules concerning this sub-task:
• Cor is the correct calculation of the sum a + b (e.g., 2x − 5x 7→ −3x or
2x3 + 5x3 7→ 7x3 ).
• PlusOp, meaning PlusOpposite, is the sum of a and the opposite of b. The
rule can be written as a+b 7→ a−b (e.g., 2x − 5x 7→ 7x or 2x+5x 7→ −3x).
• OpPlus, meaning OppositePlus, is the sum of the opposite of a and b.
The rule can be written as a + b 7→ −a + b (e.g., 2x − 5x 7→ −7x or
2x + 5x 7→ 3x).
• OpPlusOp, meaning OppositePlusOpposite, is the sum of the opposite of
a and the opposite of b. The rule can be written as a + b 7→ −a − b
(e.g., 2x − 5x 7→ −3x or 2x + 5x 7→ −7x).

The Contexts.
We decided to restrict the context variables associated to this task. We chose
only two variables: sign of a and b, order of |a| and |b|.6
The first variable can be decomposed into four binary variables: (sign of
a is plus, sign of b is plus), (sign of a is minus, sign of b is plus), (sign of a
is plus, sign of b is minus) or (sign of a is minus, sign of b is minus). The
second variable can be decomposed into three binary variables: |a| is smaller
than |b|, |a| is greater than |b| or |a| is equal to |b|.
For example, in the expression 2x − 5x, the sign of a is plus and of b
is minus. We denote this information by P M (“Plus, Minus”). In this same
expression, |a| is smaller than |b|. This information is denoted C1. The binary
variable |a| is greater than |b| is denoted C2, and the variable a is equal to b
is denoted Eg.
Therefore, there are 12 possible contexts for the task of collecting like terms
denoted by juxtaposing the two binary variables. For example, the context
C1P M means that the context is (sign of a is plus, sign of b is minus, |a| is
smaller than |b|).

The Attributes:
Ordered pair (Context, Rule). Since there are twelve contexts and four rules,
there are 48 attributes. Their name is obtained by the association of the rule
name and the context name. For example, the attribute CorC1PM means
the use of a correct rule, Cor, in the context C1P M , while the attribute
PlusOpC2MP means the use of the rule PlusOp in the context C2M P .
6
We are aware that we do not take into account some important variables such
as the degree of the monomial, a presence of a minus sign in the expression of
the student’s step, or the nature of the coefficients. One of the reasons for this
restriction is that the automation of the process can be long and the work is still
in process: we want to be sure that we obtain results before continuing coding
other variables.
94 M.-C. Croset et al.

The Experiment.

As was mentioned in the section 7, experiments were designed with the pur-
pose of obtaining data about collecting like terms. However, this task appears
in most of exercises that were proposed in these experiments. In this exam-
ple, the students population consists of a group of three grade 8 classes of 87
students in total (n = 87). The table is then an 87 × 48 matrix.
Each student worked on more or less 28 exercises. In total, 2584 elementary
steps concern the chosen sub-task. We obtain the average of 29.7 elementary
steps per student.

Some Results.

In what follows, we describe five of the behaviour groups that are highlighted
by the treatment of a similarity tree, see Fig. 6.
The behaviour class (OpP lusOpC2M M , OpP lusOpC1M M ) can be in-
terpreted as follow: when confronted with a sum of two negative terms, a and
b, the students from this behaviour class sum |a| and |b|. For example, they
transform the expression −2x−5x into 7x or −5−2 into 7. This can mean that
they understand that they have to add |a| and |b|, but they make a mistake
for the sign of the result, making it always positive.
The second behaviour class (CorC1P P , CorC1M P , OpP lusOpC2M P ,
OpPlusOpC1PM, OpP lusOpC2M M , OpP lusOpC1M M ) looks more pre-
cisely at the meaning of the previous behaviour class. Let us take the example
of the six attributes of this class given in Tab. 5.

Attribute Example Example name

CorC1P P 2 + 5 7→ 7 T1
CorC1M P −2 + 5 7→ 3 T2
OpP lusOpC2M P −5 + 2 7→ 3 T3
OpP lusOpC1P M 2 − 5 7→ 3 T4
OpP lusOpC2M M −2 − 5 7→ 7 T5
OpP lusOpC1M M −5 − 2 7→ 7 T6

Table 5. Example of the six attributes of the class (CorC1P P , CorC1M P ,

OpP lusOpC2M P , OpP lusOpC1P M , OpP lusOpC2M M , OpP lusOpC1M M )

These students sum correctly when they are confronted with a sum of two
positive terms, or with a sum of a negative term (−5), smaller (in absolute
value) than a positive term (2), (T 1, T 2). When they are confronted with
a sum of a negative term greater (in absolute value) than a positive term,
the negative term being on the left (T 3) or on the right (T 4), the students
calculate | − 5 + 2|. The cases T 5 and T 6 have been already explained in the
previous class.
 Student’s Algebraic Knowledge Modelling 95

We can interpret this group as follow: the students know that for a sum of
one positive and one negative terms, they have to subtract the smaller term
from the greater one; however, they do not know that the result must have
the same sign as the greater term. For a sum of two terms with same sign,
they know they have to add the two terms. However, they do not know that
the sign of the result is the same as the sign of the terms. That is particularly
visible when the two terms are negative. What they may be doing is that
they multiply the signs of the two terms, as if it were a multiplication. They
apply the rule: ∆|a| + ∆|b| 7→ |a| + |b|, where ∆ is the sign of the term a
and b.7 This interpretation was verified by replaying the work of the students
who contributed the most to this class, via the ‘replay system’ in Aplusix. In
particular, we observed that these students are never wrong about the sign
in the case of a multiplication of two terms: they mastered the sign rules for
multiplication. In order to verify this hypothesis it is possible to insert rules
about multiplication of two terms in the CHIC table and see whether such
groups emerge.
The third behaviour class (CorC2MP, CorEgMM, CorC1MM, CorC2MM )
is a class of correct actions.
The fourth & fifth ones (PlusOpC1PP, OpPlusC2PP, OpEgPP, OpPlus-
C1PP, PlusOpC1PP ) seem strange: the students subtract one term from the
another even if neither of them has minus sign: the contexts are all positive
terms (P P ). This is probably due to the choice of the variables: we have
selected only two context variables. If we consider and add the variable “pres-
ence of a minus sign in the source of the student’s step”, we would perhaps
have an explanation for this class. Behaviour classes such as those generalising
behaviours of the students A and B presented in section 4 may appear.
2MP

OpC PM
1MM

2MM
1

P
P

P
OpC
OpP usOpC

OpP OpC1P
Cor usOpC

C2P

C1P

C2P
P
Cor 1PP

Cor 2PM

M
P

M
lus
C1M

lus

EgP

EgM

C2M

lus

gPP
lus

sOp
C1M

C1P

C2M

C1M
C

s
l

C
OpP
Cor

OpP

Plu
OpP

Cor

Cor

Cor

Cor

OpE
OpP

Plu
Cor

Cor

Fig. 6. Part of the similarity tree

We obtained a collection of ordered pairs (context, rules) that are repre-

sentative of stable behaviours for a part of the students population. It was
possible to explain these behaviour groups cognitively. We could thus reach a

7
Application of an incorrect rule leads to a correct result, in the case of the sum
of two positive terms.
96 M.-C. Croset et al.

higher level description of behaviours than in the previous sections. A direct

application of these results would be modelling didactical decisions in terms of
mapping didactical explanations to the identified behaviour groups, in order
to address to each student an appropriate and easily understandable message.

9 Conclusion

Errors, and more globally behaviours in which we are the most interested,
are those that are “persistent and reproducible”. Catching stable errors and
characterizing types of these errors could allow teachers, didacticians and
artificial tutors to take adequate and personalized didactical decisions instead
of using a systematic and repetitive feedback.
Data on which our stability research is based are collected in the learning
environment for algebra, Aplusix. Each student’s action is viewed as a rule
application. Our purpose was to determine rules that are used regularly by
the student and to link them to the algebraic context in which the student
used them. We aimed at both a detailed description of the student’s learning
state and a high cognitive level of its description. However, mapping a single
action to stable behaviours is not an obvious task.
Our research work proceeded in two phases. First, we have established a
list of the sub-tasks that provoke regular errors. For that, we needed to deter-
mine a good granularity of the sub-tasks. We used the Statistical Implicative
Approach (SIA) to determine, in a students population, the variables that
cause a stable erroneous behaviour. Second, a systematic analysis of each
sub-task allowed us to outline the main behaviour groups.
Thanks to the Statistical Implicative Analysis theory, a certain stability
of behaviours has been observed by associating algebraic context variables to
actions. It has been possible to determine for a student the possible tasks and
sub-tasks that are interesting to model because they provoke stable behaviours
in the student. Moreover, the SIA has permitted to point out the algebraic
context variables that are source of actions: implicative links between variables
and student’s actions have been provided as a result of this analysis at a very
fine grain size. A statistical implicative analysis of frequencies of ordered pairs
(algebraic context, action) for a students’ population has allowed behaviour
groups to emerge: groups of ordered pairs the most used by a part of the
population or those that show a very high level of stability of frequencies.
Using a statistical approach in a didactical research is an original method:
most of the didactical research cannot afford to deal with a large student body.
In addition, interactive learning environments providing models of students
reach either a very fine level of description where steps are not linked together,
or a coarse grain size with general information not linked to a precise domain
(such as information about whether the student learned best with a directed
or explanatory learning tasks [19]).
 Student’s Algebraic Knowledge Modelling 97

This research work is based on a description of the algebraic context vari-

ables that is a deep and hard work and that has important consequences on
the modelling results. The obtained results seem to prove the relevance of
our choices: students’ actions are explained by the presence of some of these
algebraic context variables. Our work also used elementary steps provided by
an automatic diagnosis done within the learning environment. It is therefore
dependent on the quality of this diagnosis and we need to study carefully the
consequences of a bad diagnosis.
The results presented show a great didactical interest both for teachers and
artificial tutors designers. Automatic diagnosis of students’ knowledge state in
terms of correct and erroneous rules applied in particular algebraic contexts
provides the teacher with the opportunity to focus remediation to the very
source of the incorrect behaviour. The presented students’ knowledge mod-
elling opens the possibility to build a model of didactical decisions allowing
to produce an appropriate feedback as a response to students’ action.

References
1. G. Brousseau. Les obstacles épistémologiques et les problèmes en mathématiques.
Recherches en Didactique des Mathématiques, volume 4–2, pages 165–198, 1983.
2. R. Sison, M. Shimura. Student Modeling and Machine Learning. International
Journal of Artificial Intelligence in Education, volume 9, pages 128–158, 1998.
3. D. H. Sleeman, A. E. Kelly, R. Martinak, R. D. Ward, J. L. Moore. Studies
of Diagnosis and Remediation with High School Algebra Students. Cognitive
Science, volume 13, pages 551–568, 1989.
4. J. S. Brown, K. Van Lehn. Repair theory: A generative theory of bugs in proce-
dural skills. Cognitive Science, volume 4, pages 379–426, 1980.
5. R Gras. L’analyse implicative: ses bases, ses développements. Educaçao Matem-
atica Pesquisa, volume 4:2, pages 11–48, 2004.
6. C Kieran. The core of algebra: Reflections on its Main Activities. ICMI Algebra
Conference, Melbourne, Australia, pages 21–34, 2001.
7. J. R. Anderson, A. T. Corbett, K. R. Koedinger, R. Pelletier. Cognitive Tutors:
lessons learned. The journal of the learning sciences, volume 4:2, pages 167–207,
1995.
8. D. McArthur, C. Stasz, M. Zmuidzinas. Tutoring Techniques in Algebra. Cogni-
tion and Instruction, volume 7:3, pages 197–244, 1990.
9. M. Beeson. Design Principles of Mathpert: Software to support education in
algebra and calculus. In N. Kajler, editor, Computer-Human Interaction in
Symbolic Computation, Springer-Verlag, Berlin, Heidelberg, New York, pages 89–
115, 1998.
10. R. Prank, M. Issakova, D. Lepp, V. Vaiksaar. Using Action Object Input Scheme
for Better Error Diagnosis and Assessment in Expression Manipulation Tasks.
Maths, Stats and OR Network, Maths CAA Series, 2006.
11. J. F. Nicaud, D. Bouhineau, H. Chaachoua. Mixing microworld and CAS fea-
tures in building computer systems that help students learn algebra. International
Journal of Computers for Mathematical Learning, volume 9:2, 2004.
98 M.-C. Croset et al.

12. A. T. Corbett, J. R. Anderson. Knowledge Tracing: Modeling the Acquisition

of procedural knowledge. User modeling and user-adapted interaction, volume 4,
pages 253–278, 1995.
13. S. J. Payne, H. R. Squibb. Algebra Mal-Rules and Cognitive Accounts of Error.
Cognitive Science, volume 14:3, pages 445–481, 1990.
14. N. Balacheff. Les connaissances, pluralité de conceptions. Le cas des mathé-
matiques. In P. Tchounikine, editor, Actes de la conférence Ingénierie de la
connaissance, Toulouse, pages 83–90, 2000.
15. S. Soury-Lavergne, N. Balacheff. Baghera Assessment project: Designing an Hy-
brid and Emergent Educational Society. Cahier du laboratoire Leibniz, volume 81,
2003.
16. Y. Chevallard. Concepts fondamentaux de la didactique: perspectives apportées
par une approche anthropologique. Recherches en didactique des mathématiques,
La Pensée Sauvage, volume 12:1, Grenoble, pages 73–111, 1992.
17. G. Brousseau. Théorie des situations didactiques. La Pensée Sauvage, Grenoble,
1998.
18. J. F. Nicaud. Modélisation cognitive d’élèves en algèbre et construction de
stratégies d’enseignement dans un contexte technologique. Project report of the
“Ecole et sciences cognitives” research programme, Cahier du laboratoire Leibniz,
volume 123, 2005.
19. M. Quafafou, A. Mekaouche, H. S. Nwana. Multiviews learning and intelligent
tutoring systems. Proceedings of Seventh World Conference on Artificial Intelli-
gence in Education, volume 7, 1995.
The graphic illusion of high school students

Eduardo Lacasta and Miguel R. Wilhelmi

Departamento de Matemáticas, Universidad Pública de Navarra

31006 Pamplona (Navarra), Spain
{elacasta, miguelr.wilhelmi}@unavarra.es

Summary. The factorial analysis of the relationship between the mathematical

background on linear and quadratic functions, on the one hand, and the representa-
tion of functions (graphics, figures and so on) on the other hand, stands in contra-
diction to the usual assumption of the existence of a “graphical conceptualization” of
functions, different from the “non-graphical conceptualization”. Nevertheless, both
the authors of scholar texts and the teachers involved in this research tend to use
the graphical representation of functions. In the context of proportionality, the Sta-
tistical Implicative Analysis of students’ preferences regarding the kind of graphical
representation reveals the existence of a graphical illusion shared by high school
students.

Key words: Mathematics Education, function, graphics, Statistical Factorial

Analysis, Statistical Implicative Analysis.

1 The function and its graphic: its basis and its form
In current teaching there is a trend that favors the presentation of mathe-
matical notions (notably that of functions) in a visual way. This trend relies
on curricular developments and in changes of teaching methods based on
the growth and development of information and communication technologies
(ICT), such as graphic, symbolic and programmable calculators or mathe-
matical software as CABRI, Derive, Maple or Mathematica. These curricular
developments and these changes in teaching are in many cases influenced by
the Principles and Standards for School Mathematics [9].
“[Technology Principle] Electronic technologies (calculators and com-
puters) are essential tools for teaching, learning, and doing math-
ematics. They furnish visual images of mathematical ideas [. . . ]
Instructional programs from pre-kindergarten through grade 12
should enable all students to understand patterns, relations, and
functions [Standard Algebra]. In grades 6–8 all students should:
E. Lacasta and M.R. Wilhelmi: The graphic illusion of high school students, Studies in
Computational Intelligence (SCI) 127, 99–117 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
100 Eduardo Lacasta and Miguel R. Wilhelmi

• represent, analyze, and generalize a variety of patterns with

tables, graphs, words, and, when possible, symbolic rules;
• relate and compare different forms of representation for a rela-
tionship;
• identify functions as linear or nonlinear and contrast their prop-
erties from tables, graphs, or equations [14]”.
In school, the introduction and the development of the notion of function
often observe an intuitive approach based nearly exclusively upon graphic
language. This teaching practice may have the following risk: “to only teach
properties of the functions that are specific to the graphic context” [5]. There-
fore it is necessary to analyze the priority of graphic language in the teaching
of Mathematics.
Numerous studies in Mathematics Education are theoretically based and
experimentally supported by the importance of visualisation in teaching
([3, 6–8, 10, 16, 18], etc.). A naïve or hasty application of these studies may
excessively emphasize this intuitive and graphic approach of the notion of
function, with the illusion that the students will have the ability to identify
and represent the same concept in different representations, and the flexibility
in moving from one representation to another. Therefore, these applications
ignore the phenomenon of compartmentalization [6].
All this has lead us to study the role of cartesian graph of functions (CGF)
in Secondary teaching. We will examine students according to their compe-
tence in the resolution of presented problems: a) in an exclusively textual way,
b) with a numeric table or c) by means of a graphic. Also, the relationship of
those competences with the preferences declared by the students (for the way
of presentation of the functions) is examined.
Does some mathematical competence include the other competences? Is a
graphically competent student also competent in textually presented problem
solving and in table presented problems? What role does the Cartesian graph
of functions play in problem solving, especially in the case of relationships
between linear functions and proportionality?
Lacasta identifies and describes five different functions of CGF like “ma-
terial context (milieu matériel )” [12]. This complexity of CGF suggests that
it is necessary to redefine the role attributed to GCF in Secondary school. It
is also necessary to relate this material context to other forms of representing
functions, such as the numeric table and the text.
In fact, the object-representation dichotomy is problematic. Font, Godino
and D’Amore justify that each object-representation pairs (without segre-
gation) permits a subset of practices of the whole set of practices that are
considered the unique and holistic meaning of the object [7]. However, in
each subset of practices, the object-representation pair (without segregation)
is different, in that it makes different practices possible.
The objective of this work is to contrast the following hypotheses:
 The graphic illusion of high school students 101

[H1] There is no relationship of statistical inclusion amongst the sets of stu-

dents defined by the three competences (textual, numeric or graphic). In
figure 1 a graphic representation of this hypothesis is shown.
This hypothesis expresses the idea that competent students exist espe-
cially in the resolution of problems given in a certain presentation way.
But the field of the each competence —Ctx, Ctb, Cg— would be no more
than an approach to the global mathematical competence, with apprecia-
ble areas of failure.
[H2] Students prefer a given form of presentation because it improves their
competence in the outlined problems (figure 1).
This hypothesis expresses the idea that the perception about the difficulty
of a given question (following the way of presentation) is an indicator
of efficiency in problem solving: The students are aware of their real
capacities.

Caption
Cpsf: competence in problems solving involving linear and quadratic functions
Ctx: competence in textually presented problems solving
Ctb: competence in problems solving, when the function is represented by
a numerical table
Cg: competence in problems solving, when the function is graphically represented
Ptx: student’s preference for the textual presentation of the function
Ptb: student’s preference for the tabular presentation of the function
Pg: student’s preference for the graphical presentation of the function

Fig. 1. Graphic representation of hypotheses 1 and 2

102 Eduardo Lacasta and Miguel R. Wilhelmi

The contrast of the hypotheses will give us information of the running of

the didactic system, this is, information of the mechanisms that determine the
construction and communication processes of knowledge relative to functions
in high school. However, we have shown how the implication between two vari-
ables (competence or attitudinal) doesn’t necessarily represent a cause-effect
relationship and, therefore, it doesn’t allow to establish teaching guidelines.
To contrast these hypotheses two questionnaires were given to a sample
of 87 students. The description of the answers to the first questionnaire was
carried out by means of a factorial analysis (Factorial Analysis of Correspon-
dences and Main Components Analysis). The description of the answers to
the second questionnaire was carried out by means of a Statistical Implicative
Analysis (SIA).
A more general goal of this paper is to show how:
1. The SIA contributs to validation or refutation of hypotheses.
2. The SIA can be integrated with other statistical analyses (FAC and MCA,
in particular).
3. The SIA yields conclusions that cannot be obtained with other forms
of analysis by allowing the contrast between a priori and a posteriori
analysis.
Shortly, in this paper we apply the SIA to a notable problem in mathe-
matics didactic (the role of representations in the learning of mathematics)
and we establish certain fundamental aspects of the SIA too.

2 A priori analysis and factorial analysis

To determine the existence of a graphic way of solving problems, we gave 87

Spanish Secondary school students a first questionnaire (see Appendix). The
mathematical knowledge proposed was the following: the reading of intersec-
tions, the sign of functions for the intervals of variable x, the comparison of
functions, the extrapolation, etc. We have limited our study to the use of
polynomial functions of first and second degrees, which are known by all the
students in the sample.
The functions in this questionnaire have been represented either by means
of a Cartesian graph (G), a numerical table (T) or an algebraic formula (F).
We have defined variables according to the knowledge proposed and its way
of presentation. Each knowledge has been proposed following the three ways
of presentation: G, T and F. In table 1 the set of defined variables is shown.
An a priori analysis establishes a model according to which the students’
competences vary depending on the three kinds of presentation G, T and F.
However, the results of the questionnaire show that the examined competences
are not grouped following the a priori model.
 The graphic illusion of high school students 103

its sig com reg ext max Cp Cd

G 1G 2G 3G 4G 5G IpG CpG CdG
F 1F 2F 3F 4F 5F IpF CpF CdF
T 1T 2T 3T 4T 5T IpT CpT CdT

Caption
The way of presentation and the knowledge concerned for every question is
determined respectively by margins of line and column
F: algebraic formula representation
T: numerical table representation
its: reading of the intersections
sig: sign of the functions for intervals of variable x
com: comparison of functions
rég: partition of the plane by functions.
ext: extrapolation
max: maxima and minima
The “max” column corresponds to questions on the identification of maxima on the
quadratic function (identification on the parabola “Ip”)
The “Cp” column corresponds to the calculation of increasing/decreasing intervals
of the parabola
The “Cd” column corresponds to the determination of the increasing/decreasing
character of the line
The numbers correspond to the groups of questions, from 1 to 5

Table 1. Variables of questionnaire 1

The empirical data show a gap between the a priori model and the pro-
cedures really observed. The used method was the factorial analysis (Fac-
torial Analysis of Correspondence —FAC— and Main Components Analysis
—MCA), since it was about “revealing symmetrical relationship and establish-
ing discriminative factors in a population by the way of variables”. In figures
2 and 3 the first two main factors are shown.
In the planes of the two main first factors (figures 2 and 3), the pupils’
success to questions appear grouped according to the mathematical problem
and not according to the kinds of presentation of the function (G, F or T).
The kind of presentation of the questions does not influence the global
success of students in the questionnaire. Thus, there is no sign of the exis-
tence of the graphic conception [1] different from a non graphic conception of
mathematical notions. Contrary wise, the student’s responses vary according
to mathematical notions and not according to the way it is presented (G, T
or F).
104 Eduardo Lacasta and Miguel R. Wilhelmi

Fig. 2. First questionnai1.8 FAC plan of the success matrix

In the present conditions of math teaching in Secondary, the graphs1 don’t

seem to play a crucial role in the learning of functions. Evidently, the graph on
its own cannot support all mathematical knowledge that it represents, espe-
cially in the case of relationships between linear functions and proportionality.

1
The concerned graph is always the Cartesian graph. The function is a character-
istic notion of the Secondary education. In Primary education the influence of
the graphic presentation of other mathematical notions can condition the ability
of pupils in problem solving [6].
 The graphic illusion of high school students 105

Fig. 3. First questionnai1.8 MCA plan of the success matrix

Pantziara, Gagatsis and Pitta-Pantazi provide an evidence of external va-

lidity of our conclusions through inferencial statistic (t-test) [15]. These au-
thors reach a similar conclusion: “The results of the study suggest that the
presence of the diagrams did not increase students’ ability in solving the non
routine problems [. . . ] The results of the study show that the efficient use of
a diagram did not imply the successful solution of a problem and reversely
106 Eduardo Lacasta and Miguel R. Wilhelmi

the successful solution of a problem did not imply the efficient use of the
accompanied diagram” [15, p. 495].

3 Ostensive use of the graph in teaching

Teaching cannot do without the resource to show a part for the whole, this
is, to identify by means of examples the generic mathematical objects. The
ostention is the following didactic phenomenon: the teacher that has shown a
copy of, for example, an increasing function, has the illusion that he has really
communicated the mathematical notion of an increasing function.
“We postulate the appearance of didactic phenomena [. . . ] on the dif-
ficulties of generating [. . . ] ostensive instruments, necessary for
the progress of the mathematical activity, and frequently valued
culturally by the school, but which cannot receive a clear mathe-
matical status. This contradiction between, on one hand, the phe-
nomenon of the ostensive reduction and, on the other hand, the
cultural valuation of the ostensive instruments, considered indis-
pensable for a ‘significant’ mathematical activity, doesn’t seem to
be able to be solved, in the usual didactical contracts, but by means
of a certain mathematical activity” [2, p. 106].
Some time ago [11], we verified that most of high school teachers prefer,
in the teaching of the functions, the conditions that better allow an ostensive
didactic contract [4]; that is to say, high school teachers prefer the graphic
representation of functions in the measure that it favours an ostensive con-
tract.
The importance assigned by teachers to the Cartesian graph is based on a
“false transparency” attributed to this graph. In other words, the represented
function would be directly “visible” on the Cartesian graph that represents it.
The didactical phenomenon of the “false transparency” can be described by
the following computer science metaphor: Cartesian graph is for the teacher
an editor WYTIWYG (What You ‘Think’ Is What You Get), whereas for the
student in many cases it is an editor WYSIWYG (What You ‘See’ Is What
You Get).
The evidence illusion [12] is a more general phenomenon: the professor
observes the notion that he wants to teach in the representation of an object,
while the student doesn’t transcend to the representation; this is, the student
sees the representation like a mere representation (“as such”).
This vision of the teachers is based on the following belief (implicitly ac-
cepted): to achieve the learning of complex analytic notions, it is necessary
to use a presentation that is easy for the students to learn; that would be the
graph. Said otherwise, the graphical context facilitates the learning of complex
notions. “[. . . ] But many students cannot utilize their visual representations
to advance in their problem solving” [17, p. 315].
 The graphic illusion of high school students 107

4 Similarity Analysis and Implicative Analysis

The first questionnaire is a general and exploratory analysis. We need empir-

ical data to analyze the observed facts in the case of relationships between
linear functions and proportionality. For it, we need a new questionnaire and
a specific statistical method to achieve non symmetrical relationships among
curricular, attitudinal and competence variables (SIA). In fact, the application
of the SIA to the first questionnaire doesn’t determine additional information
of the achieved by the precedent analyses (FAC and MCA).
A second questionnaire on the notion of proportionality was given to stu-
dents. The students had to solve the same kind of mathematical problems:
given four pairs of values, they had to determine whether the respective mag-
nitudes were proportional. The enunciation of the problems was given in three
different ways: G, T or F. The students had to say if the magnitudes were
proportional or not and justify their answers.
Before carrying out these tasks, some students had been trained on the
graphic representation of functions but others had not. All of them stated
their preferences on the way of presentation of the tasks: textual, through a
numeric table or by means of graphics.
The second questionnaire searches contrasting hypotheses. The empirical
master plan must take into account several variables, susceptible of having a
previous control of their impact on the contingent data. The types of variables
are: problem variables (modes of presentation of the problems, the formula-
tions, type of relation between problem variables) and an individual variable
(type of classes to which the students belong). A Chi-square test shows that
student’s responses are similar in different types of observation according to
problem variables. This fact allows us to define some dichotomic variables
(RTx, RTb, RG). We assign the value 1 to a variable if an individual satis-
factorily carries out at least 70% of the questions associated to this variable.
Otherwise, we assign value 0.
Based on these facts (type of received training and stated preferences)
and the answers of the students to the questionnaire we defined the following
variables:
• A curricular variable (Erf): training to the representation of functions.
• Three attitudinal variables (PTx, PTb, PG): preferences to the different
kinds of presentation.
• Five competence variables (Da, RG, REU, RTx, RTb): Students success
to answer the questions based on type of presentation.
All the variables are binary (0, absence; 1, presence). Frequencies and
percentages of variables appear in table 2.
The Similarity Analysis [13] shows that the preference of presentation
with tables or text is related to the students’ success in problem-solving pro-
posed by means of numeric tables (figure 4). Nevertheless, the preference for
graphic representation is only linked to the numerical resolution of graphically
108 Eduardo Lacasta and Miguel R. Wilhelmi

Variable RTx RTb RG REU Ngi Erf Da PTx PTb PG

Frecuency 62 72 46 67 61 54 17 18 22 28
Percent (%) 71.27 82.76 52.88 77.02 70.12 62.07 19.55 20.69 25.29 32.19

Table 2. Frequencies and percentages of variables

proposed questions. But this preference is neither associated to graphic rep-

resentation training nor to the students’ success in the resolution of graphic
problems.

Caption
Ngi: numerical resolution when the problem is presented by an incomplete
graph
RG: success in graphically presented problems
RTx: success in textually presented problems
RTb: success in numerical table presented problems
REU: global success in the questionnaire
PTx: preference for textual presentation
PTb: preference for tabular presentation
PG: preference for graphical presentation

Fig. 4. Similarity graph

Table 2 reveals that students encounter greater difficulties in graphically

presented problems than in textually or in numerical table presented problems.
In turn, the textual tasks are more difficult than the tasks involving tabular
information. Moreover, in the similarity graph the connections of students’
success at the graphic tasks, at the textual tasks and at the tabular tasks are
relatively weak. These remarks suggest that some kind of compartmentaliza-
tion in students’ performance across the various representations exists.
 The graphic illusion of high school students 109

The success in the resolution of graphic problems is linked to the technique

of determination of the proportionality by means of the alignment of the points
with the origin in the graphic. The training on representation of functions is
linked to this success and this technique too.
REU comprehends competences RTx, RTb and RG. The similarity graph
shows that the contribution of the competences RTx and RG to REU are more
outstanding than the competence RTb.
The implicative graph (figure 5) shows that the set of students stating their
preference for graphic language is not included (not even quasi-included) sta-
tistically in the set of successful students in the graphical tasks: that is, the
graphic preference (PG) is implicatively isolated; the bigger implicative index
involving PG is 53 (PG−→Ngi, see table 3). Furthermore, there is no impli-
cation in any direction between graphical training (Erf) and global success in
the questionnaire (REU).

Variables RTx RTb RG REU Ngi Erf Da PTx PTb PG

RTx 0 91 99 0 0 0 0 0 0 0
RTb 0 0 0 0 0 0 0 0 0 0
RG 99 97 0 100 75 97 0 0 0 0
REU 0 98 0 0 0 0 0 0 0 0
Ngi 68 59 0 90 0 0 0 0 0 0
Erf 60 72 0 88 75 0 0 0 0 0
Da 67 0 100 87 0 100 0 58 61 0
PTx 72 94 72 73 0 0 0 0 91 0
PTb 0 85 0 50 0 0 0 0 0 0
PG 0 0 0 0 53 0 0 0 0 0

Table 3. Implicative index

The results obtained through the Statistical Implicative Analysis (SIA)

determine an evidence of the existence of a graphic illusion of secondary stu-
dents.
The Implicative Analysis shows that the set of students defined by graphic
success is included statistically in the sets defined by the textual and numeric
success (figure 5). The success in the graphic resolution is only achieved by
students able to also solve the problems presented by the means of a numerical
table and by the means of text. In other words, only some competent students
in the resolution on tabular and textual problems achieve the graphical com-
petence.
These results refute hypothesis H1. Against what was affirmed in H1, the
empirical data of the behaviours exhibited in the given sample establish a
statistical inclusion relationship between the mathematical competences and
the method of representation used. In fact, students that solved the problems
110 Eduardo Lacasta and Miguel R. Wilhelmi

Caption
Dashed arrows: transitivity of the statistical implication
Continuous arrows: implication (99%, 95%, 90%)
Vertical scale: ratio of students giving evidence of the respective variables

Fig. 5. Implicative graph

presented graphically were in general able to solve these problems presented

in a tabular and textual way, but not vice versa.
Hypothesis H1 suggests the existence of three independent competences to
(numerical table, textual and graphical) that would be different approaches to
a same mathematical knowledge. There would exist three different conceptions
for these mathematical notions. However, the refutation of H1 comforts the
thesis formulated in the a priori analysis (section 2): students’ responses vary
according to mathematical notions and not according to the way it is presented
(G, T or F).
The refutation of H1 also questions the following extended belief among
high school teachers: students learn more easily the notions of functions using
graphic representation.
On the other hand, the implicative graph (figure 5) refutes hypothesis
H2 partially. Indeed, the sets of students that prefer the tabular or textual
presented problems are included in the set of the students that solve those
problems satisfactorily. However, the set of students that prefer the graphic
presentation is not included statistically in the set of students that solve the
graphical represented problems.
 The graphic illusion of high school students 111

5 Synthesis and conclusions

We used a sample of 87 Spanish High School students (13 years old). The
functions in a first questionnaire have been presented either by means of a
Cartesian graph (G), a numerical table (T) or an algebraic formula (F). The
factorial analysis has allowed to explore the data and it has detected the
statistical proximity (symmetrical relationship) of the behaviours, according
to the treated notions. Indeed, the proximity doesn’t obey the form of the
tasks (graphic or non graphic presentation), but to the involved notions.
In a second questionnaire, on one hand, the proximity of some variables
is detected in the Similarity Analysis. This way, the relationship between the
variable “instruction in graphic representation” and the variable “detection of
the proportion by the way of a graphical technique” is a very strong one. In ad-
dition, the relationship between the “preference for the tabular presentation”
and the “success in the resolution of tabular problems” is notable too.
On the other hand, the SIA detects statistically significant relationships
between the preference for textual presentation and the success in textually
presented problems, that the Similarity Analysis does not show. We also ob-
serve that the preference for the graphic presentation doesn’t imply the success
in its resolution.
These facts take us to conclude that a graphic illusion exists amongst
these students, since similarity relationships do not exist (neither implica-
tive relationship) between the preference for the graphic presentation and the
competence in the resolution of problems presented graphically.
The SIA contributes new elements, even though of delicate interpretation.
The empirical data doesn’t allow to justify the application of a specific teach-
ing method. In fact, there is a radical rupture between explanatory didactics
and normative (or technique) didactics based on two facts:
1. The statistical implication is not transitive. The competence in the deter-
mination of the proportionality by means of the alignment of the points
in the graphic with the origin (Da) implies the success in graphically
presented problems (RG). The competence RG implies the success in tex-
tually presented problems (RTx). However, it’s not possible to deduce
that “Da” implies “RTx”. Consequently, it is not possible to establish the
maxim: “If the students are competent in ‘Da’, then most of them solve
the presented problems in a textual way (RTx)”.
2. The statistical implication admits an interpretation (mathematically iso-
morphic) according to the Set Theory 2 . Given a set E, we defined P (Q
respectively) like the set of elements x in E (x ∈ E) that verify the propo-
sition p (q respectively):

P = {x ∈ E | p(x)} (Q = {x ∈ E | q(x)} respectively)

2
The Boolean algebra rules can be written in both set and logic notation.
112 Eduardo Lacasta and Miguel R. Wilhelmi

Then “p implies q” is equivalent to affirm that P ⊆ Q. This fact explains

that the success in graphically presented problems (RG) implies graphical
training (Erf). Only a subset of students who have received graphical
training have succeeded in graphically presented problems (RG implies
Erf).
3. The statistical implication doesn’t necessarily determine a cause-effect re-
lationship. The success in graphically presented problems implies the suc-
cess in the textually or tabullary presented problems. This fact doesn’t
determine a preferential order in teaching for the different kinds of repre-
sentation. There is no cause-effect relationship either between the resolu-
tion of graphic problems and that of other types of problems.
The results in SIA are static, that is, they represent a moment of the school
reality. The fact “p implies q” doesn’t mean that “p happens before q”, neither
that “p causes q”. In this case, the conclusion is that only some competent
students in tabular and textual contexts reach the graphic competence.
In Mathematics Education, the statistical analysis of empirical data doesn’t
explain the running of didactical systems; neither does it determine strategies
for intervention and control. However, the statistical methods allow to con-
front the a priori analysis with the a posteriori analysis. Also, the SIA finds
conclusions non detectable from other statistical methods; these conclusions
allow to interpret more efficiently the didactical systems and to value the
strategies designed for the intervention in these systems.
Some students perceive the graphic representation as an easy and intuitive
instrument of reaching the mathematical knowledge, but this perception is
contradicted. What role do the educational practices of high school teachers
play in this perception? Studies are needed that would allow to establish
normative approaches for teaching improvements of functions in high school
and for training improvements of those teachers.

References
1. N. Balacheff. Conception, connaissance et concept. Séminaire DidaTech, Uni-
versité Joseph Fourier (Paris), 1995.
2. M. Bosch and Y. Chevallard. La sensibilité de l’activité mathématique aux
ostensifs. objet d’étude et problématique. RDM, pages 77–124, 1999.
3. E. G. Bremigan. An analysis of diagram modification and construction in stu-
dents’ solutions to applied calculus problems. JRME, pages 248–277, 2005.
4. G. Brousseau. Theory of didactical situations in mathematics. Kluwer Academic
Publishers, 1997.
5. G. Chauvat. Courbes et fonctions au college. Px, pages 23–44, 1999.
6. I. Elia, A. Gagatsis, and R. Gras. Can we ‘trace’ the phenomenon of com-
partmentalization by using the implicative statistical method of analysis? an
application for the concept of function. In R. Gras, F. Spagnolo, and J. David,
 The graphic illusion of high school students 113

editors, Troisième Rencontre Internationale A.S.I. Analyse Statistique Implica-

tive. Quaderni di Ricerca In Didattica of G.R.I.M., pages 175–185 (Supplemento
2(15)), 2005.
7. V. Font, J. D. Godino, and B. D’Amore. An ontosemiotic approach to repre-
sentations in mathematics education. FLM, pages 2–7, 2007.
8. G. A. Goldin. Representational systems, learning and problemen solving in
mathematics. JMB, pages 137–165, 1998.
9. K.J. Graham and F. Fennell. Principles and standards for school mathematics
and teacher education: Preparing and empowering teachers. SSM, pages
319–328, 2001.
10. T. Graham and J. Sharp. An investigation into able students’ understanding of
motion graphs. TMA, pages 128–135, 1999.
11. E. Lacasta. Les graphiques cartésiens de fonctions dans l’enseignement sec-
ondaire des mathématiques: illusions et contrôles. Université Bordeaux 1, 1995.
12. E. Lacasta. Sur la théorie des situations didactiques, chapter Les modes de
fonctionnement du graphique cartésien de fonctions comme milieu, pages
249–262. La Pensée Sauvage, Gronoble, 2005.
13. I.C Lerman. Classification et analyse ordinale des données. Dunod, 1981.
14. NCTM. Principles and Standards for School Mathematics. E-version available
in: [http://standards.nctm.org/]. Autor, 2000.
15. M. Pantziara, A. Gagatsis, and D. Pitta-Pantazi. The use of diagrams in solving
non routine problems. In R. Gras, F. Spagnolo, and J. David, editors, Proceed-
ings of the 28th Conference of the International Group for the Psychology of
Mathematics Education, pages 489–496 (Vol. 3), 2004.
16. T. Romberg, E. Fennema, and T. Carpenter. Integrating research on the graph-
ical representation of functions. Lawrence Erlbaum Associates, Inc., 1993.
17. D. A. Stylianou. On the interaction of visualization and analysis: the negotia-
tion of a visual representation in expert problem solving. JMB, pages 303–317,
2002.
18. W. Zimmerman and S. (Eds.) Cunningham. Visualization in teaching and learn-
ing mathematics. Mathematical Association of America, 1993.
114 Eduardo Lacasta and Miguel R. Wilhelmi

Appendix: first and second questionnaires

First questionnaire

The polynomial functions f , g, h, F1 , G1 , H1 , F2 , G2 and H2 are defined for

any real number and are of degrees 1 or 2.
[A] The graphs for f , g and h are:

[B] The formulas for F1 , G1 and H1 are: F1 (x) = x2 + 2x − 3; G1 (x) = x + 3;

H1 (x) = 1 − x
[C] The numerical tables for F2 , G2 and H2 for some values of x are:

x −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7
F2 (x) −1 0 1 2 3 4 5 6 7 8 9 10 11 12
G2 (x) −21 −12 −5 0 3 4 3 0 −5 −12 −21 −32 −45 −60
H2 (x) −9 −7 −5 −3 −1 1 3 5 7 9 11 13 15 17
1. Find the values of x satisfying:
a) f (x) = h(x)
b) F1 (x) = H1 (x)
c) G2 (x) = H2 (x)
2. Are the following statements true or false?
a) If x < 0, then f (x) < 0, g(x) < 0 and h(x) > 0
b) If x > 1, then F1 (x) < 0, G1 (x) > 0 and H1 (x) < 0
c) If x > 1, then F2 (x) > 0, G2 (x) < 0 and H2 (x) > 0
3. For what intervals of x are the following inequalities true?
a) f (x) > h(x) > g(x)
 The graphic illusion of high school students 115

b) G1 (x) > H1 (x) > F1 (x)

c) F2 (x) > H2 (x) > G2 (x)
4. Find at least a pair of numbers (x, y) satisfying the following conditions:
a) y < f (x), y > g(x), y < h(x)
b) y > F1 (x), y < G1 (x), y < H1 (x)
c) y < F2 (x), y > G2 (x), y > H2 (x)
5. Starting from x = 6 (that is to say, for x > 6), which of the following
inequalities or statements are always true?
a) f (x) > g(x) > h(x)
b) F1 (x) > G1 (x) > H1 (x)
c) F2 (x) > G2 (x) > H2 (x)
d) It is not possible to assure none of them, because they can change for
further values that 6.
6. Complete the following table, putting “yes” or “not” in each box:
Graph ∃ max or min Curve Straight If: Increasing Decreasing
f x<2
x>2
g x < 1.5
x > 1.5
h x<4
x>4
F1 x < −2
x > −2
G1 x < −3
x > −3
H1 x<1
x>1
F2 x < −5
x > −5
G2 x < −1
x > −1
H2 x < −1.5
x > −1.5

Second questionnaire

1. In a stationary store 4 packages of notebooks are sold. Package A has 3

notebooks and it costs 225 euro cent (ec.); package B has 5 notebooks
and it costs 375 ec.; package C has 10 notebooks and it costs 750 ec.; and
package D has 15 notebooks and it cost 1125 ec. Is there a reduction for
buying more notebooks? Why?
2. Tests of the French High Speed train are being made in a long, plain and
straight line portion of railway. The time to travel several distances has
been timed. The results are given in the following table:
116 Eduardo Lacasta and Miguel R. Wilhelmi

Time (minutes): 3 5 9 12
Distance (Km.): 18 30 54 72
Based on the obtained results, does the train maintain its speed? Why?
3. In the Post Office, we are informed that to send packages to a same desti-
nation, the prices are according to the package’s weight and are given by
the following graph:

Is the shipment similarly expensive in all the cases or is there some vari-
ation according to the package’s weight? Why?
4. In a warehouse there are some closed packages that contain cups of the
same type. These packages have labels with the number of cups and the
total weight. This data is represented in the following graph:

Are the labels correct or has there been some error? Why?
5. In another stationary store 4 packages of notebooks are also sold. The
number of notebooks that each package contains and its corresponding
price is given in the following table:
Time (minutes): 3 5 8 20
Distance (Km.): 225 375 675 1 350
Is there a reduction when buying more notebooks? Why?
 The graphic illusion of high school students 117

6. In a freight forwarding agency we are informed that to send packages to

a same destination, for a package of 200 grs. 60 ec is paid.; for a package
of 300 grs., 90 ec.; for a package of 500 grs., 150 ec. and for a package of
600 grs., 180 ec. Represent this data graphically:

Is the shipment similarly expensive in all the cases or is there some vari-
ation according to the package’s weight? Why?
7. In another warehouse there are also some closed packages that contain
cups of the same type. These packages have some labels with the number
of cups and the total weight. On package A is placed the following label:
“3 cups, 240 grs.”. On the package B: “6 cups, 480 grs.”. On the package
C: “12 cups, 1040 grs.”. On the package D: “16 cups, 1280 grs.”. Are the
labels correct or has there been some error? Why?
8. Tests of the Spanish High Speed train are also being made in a long,
plain and straight line portion of railway. The obtained results are the
following: it takes 2 minutes in traveling 12 kilometers, 4 minutes in trav-
eling 24 kilometers, 8 minutes in traveling 42 kilometers and 10 minutes
in traveling 66 kilometers. Represent this data graphically. Does the train
maintain its speed? Why?

9. Which of these questions seemed easier to you? Why?

Implicative networks of student’s
representations of Physical Activities

Catherine-Marie Chiocca1 and Ingrid Verscheure2

1
E.N.F.A., dept CLEF, B.P. 22686, 31 326 Castanet-Tolosan, France
[email protected]
2
L.E.M.M.E., Paul Sabatier University, Bat. 3R1B2, 118 Rte de Narbonne,
31 062 Toulouse Cedex 4, France
[email protected]

Summary. The proposal reports on and discusses results of a questionnaire-based

study of young people’s attitudes (representations) to team games and volleyball
(in the context of physical education lessons). This questionnaire was given to stu-
dents in French agricultural high school. Treatment use software CHIC. Questions
approached attitudes, values and dispositions (representations) of students about
physical education, and, more particularly about volleyball. Several networks of
variables appear which make it possible to profile different kinds of students. Study
of contributions of two additional variables, sex and gender, highlighted networks,
makes it possible to improve choices of representatives networks students for later
studies based on interviews. Interestingly and somewhat unexpectedly, while sex is a
strong predictor of attitudes and dispositions to team sports and volleyball, gender
is not.

Key words: gender, sex differences, representations, sport, network

1 Introduction

We are interested in studying teaching as it is actually done and not as it could

be done. The descriptive approach to effective teaching practices is relatively
important in our work [15].
However, an earlier approach [16] to the differential dynamics of didactic
interactions according to gender in Physical Education (PE), particularly in
the case of the volleyball attack, had demonstrated the need of questioning
students about their representations of Physical and Sports Activities (PSA)
and volleyball. We call representations attitudes, values and dispositions.
We began our research by classifying students according to their represen-
tations, then studied the influence of the sex and gender variables on the
profiles that were revealed. The differences between the teachers’ interactions

C.-M. Chiocca and I. Verscheure: Implicative networks of student’s representations of Physical

Activities, Studies in Computational Intelligence (SCI) 127, 119–130 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
120 C.-M. Chiocca and I. Verscheure

and certain representative student profiles was subsequently described and

analysed.
In this paper, we have limited our discussion to the results of the ques-
tionnaire. We wanted to explore the relationships between the sex and gender
variables — as revealed by the BSRI test (the Bem Sex-Role Inventory, [3])
which was validated for physical education by [7] — and representations of
the volleyball attack.
We begin here by looking at the emergence of implicative networks that
appeared to structure the data. Then we show that the sex variable had a
greater influence than gender on students’ representations. This result leads
to a wider discussion of the validity of the BSRI test as a mean of determining
how students’ representations to PE are determined by gender.
The initial question resulted from didactic research into Physical Education re-
lating to the volleyball attack. It concerns how the didactic contract is applied
differently to males and females by PE teachers. Characteristic PE teaching
content refers to actions: changes in motor control related to ability Bouthier
and David [4] demonstrated that it is essential to take representations into
account when teaching Physical and Sports Activities in school.
We have tried to understand how students’ representations of sex and/or
gender differ with respect to PE and team sports in general and their rep-
resentations of the internal logic of volleyball in particular. In our research,
representations were regarded as variables, that we wanted to relate to the
subject variables (the students’ sex and gender), with the aim of discussing the
relationships between sex, gender and representations of volleyball, in order
to clarify our didactic approach.

2 Processing data using CHIC software program

Data was collected in clusters. A questionnaire (an extract from the ques-
tionnaire can be found in the annexe) was put to students (in scientific,
technological and professional streams) in agricultural high schools in the
Midi-Pyrénées region, in order to determine their representations of PE, team
sports and volleyball.
The questionnaire was divided into three main sections: an initial question to
establish students’ gender, based on items from the BSRI (For a discussion
of the relevance of the BSRI test for determining gender attitudes, see [15]);
a second section with questions about PE in general (its usefulness, students’
preference for a particular discipline, views on mixed-sex sports education,
teacher’s sex, etc); and a third section consisting of questions about team
sports and volleyball in particular (based on word association tests and on
semantic differenciator test inspired from Osgood, [12]).
The volleyball word association test was adapted from free-association
tests, which are verbal productions, used to study the social psychology of
representations [1]. Beginning with a starter word — “volleyball” was used
 Implicative networks of student’s representations of Physical Activities 121

here — these tests consisted in asking subjects to say any words or expressions
which came to mind (a maximum of seven here). The spontaneous nature and
the projective dimension of this production reveal the semantic universe of
the subject under study more quickly and easily than in an interview [13].
Semantic differenciator test, inspired from Osgood is a quantitative method
for analysing connotations, consisting in associating a word with pairs of op-
posing representative adjectivesx [10]. Connotation is the intensive definition
of a word. For example: the word “crow” evokes the colour black, a bad omen
and a series of implicit or explicit meanings [11]. It is a general tool for finding
meaning (words, figures, etc.) and the pairs of adjectives have to be adapted
for each case: here, for volleyball. Our work was based on research by David [5]
who used a semantic differentiator adapted for rugby to reveal the differen-
tial aspects of representations of rugby among a mixed-sex PE group. The
pairs of bipolar scales are built using antonymous adjectives drawn from word
lists. The semantic differentiator can be considered as an attitude scale for
predicting behaviour. Moreover it helps build a fairly readable image of the
representations of subject groups (particularly functional representations), an
aspect which is particularly interesting to us as far as volleyball is concerned.
These two techniques were borrowed from methodologies used to explore
students’ representations. They were chosen because they were complemen-
tary. The word association test gave the representations a “fixed” aspect,
whereas the semantic differentiator was based on tactical and dynamic ideas
of volleyball.
According to Bailleul [2] “Statistical Implicative Analysis (SIA) is a partic-
ularly effective tool for studying representations and revealing their organisa-
tional structures”. We used the CHIC software program (Cohesive Hierarchical
Implications Classification) to process the answers in the questionnaire which
related to representations of team sports and volleyball.
The CHIC software program [8, 9] performs implicative analyses, the goal
being to identify how much, statistically speaking, a particular answer to a
particular item leads to another answer to another item, thus determining the
reliability of the “quasi-implications” between variables. The software then
produces an implication diagram of the variables, leading to the identification
of networks of answers, themselves made up of “implicative chains”. The im-
plicative chains are graphic illustrations showing the implications between two
or more variables. The networks are combinations of several chains all leading
to the same variable. The various chains have similar or identical meaning
and are used to interpret the observations [2].

3 Some results
3.1 Characteristics of the population under study
At the end of 2002, questionnaires were sent out to nine General Education
and Agricultural Technology high schools in the Midi-Pyrénées region. After
122 C.-M. Chiocca and I. Verscheure

several reminders, 507 useable questionnaires were returned between Decem-

ber 2002 and February 2003. They were distributed after class by the PE
teachers themselves, to all form four students, regardless of their teaching
stream. The students were classified according to different forms of the gen-
der variable, in accordance with the BSRI method, based on the median split
method. Four forms of the gender variable were established from this: an-
drogynous (A), non-differentiated (ND), feminine (F) and masculine (M). The
gender and sex variables are independent: for instance, a female can have a
masculine or androgynous gender. The division of students according to gen-
der was quite well-balanced: 26 per cent were non-differentiated, 23 per cent
were feminine, 25 per cent were masculine and 26 per cent were androgy-
nous. In order to study the influence of the different forms of sex and gender
variables on representations of volleyball, we initialised sex and gender as “ad-
ditional variables”. “Because of their multi-dimensional nature, the implicative
analysis and the implication diagram enable us to move beyond simply ac-
knowledging the existence of a relationship between two variables, in order to
reveal meaningful oriented implicative networks” [2].

3.2 Three networks at the 0.70 threshold

At the 0.70 threshold implication level, very long chains (up to eight variables)
may be found, as well as relatively distinct networks. We shall call these
networks A, B and C. We shall begin by looking at the networks obtained
at the 0.70 threshold. Then we shall discuss these three networks further,
indicating where the different forms of the sex and gender variables had the
greatest influence on the chains.

Network A

The first network consisted of different chains involving the “team spirit” and
“I like team sports” variables. This network appeared to be characterised by
“involved” variables (resulting from both the semantic differentiator and the
word association test), all referring to the feminine characteristics of volley-
ball expressed in the literature [6, 14]: terms such as “to return it”, “don’t get
hurt”, “to train” or “to ensure”.
A certain work dimension (“to train”, technical moves (“to return it”) and
within certain limits (“don’t get hurt”) led to the development of knowing
how to “play”. This, combined with concerns that could almost be considered
as hygienically oriented (“gentle”, “to relax”, “a fun activity”) made me “feel
good”. “Feeling good”, together with the “making progress” condition (the sec-
ond work reference in this network), including the “control” elements which
enabled me to be a “team player” and “feel good” in a team activity (“I like
team sports”). This representation of volleyball (or even team sports in gen-
eral) seemed well-balanced, as it had the work dimension on one side and the
 Implicative networks of student’s representations of Physical Activities 123

Fig. 1. Network A

enjoyment dimension on the other, while it excluded the competitive dimen-

sion.
In this network, we mainly found ideas relating to progress, team spirit
and enjoying the activity. In our opinion these ideas embodied three of the
essential dimensions of volleyball’s internal logic: playing as a team, making
progress and having fun, although they overlooked the idea of it also involving
a test of strength.

Network B

With themes like “positive feelings” and “a fun activity” implying that “I like
volleyball” and that it is “a team sport”, we can hypothesise that the represen-
tations which formed this network’s structure were positive attitudes towards
this activity. Volleyball also seemed to be associated with a high regard for
physical qualities (in particular this category included words such as “tall”
and “jump high”) as well as mobility. On the other hand, the game element
seemed to be represented by “to play in continuity” (as confirmed by game-
related words such as “upwards”) and tactics.

Network C

In the third network, what predominated was the representation of the sport
as a test of team strength (“match”, “to win”, “to be ready to fight”, “play as
124 C.-M. Chiocca and I. Verscheure

Physical Positive A fun Mental

Tactics Upwards
qualities feelings activity qualities

A team To play
sport in continuity

I♥
volleyball

Body
in motion

I♥
team sports

Fig. 2. Network B

a team”), characterised by a break in the exchange, either by force (“attack”,

“rough”) or by cunning (“get clever”, “to risk”).

To become
better

To win Rough Champion

To be ready
Attack Get clever To risk Match
to fight

To control
yourself

Play as a
team

I♥
team sports

Fig. 3. Network C
 Implicative networks of student’s representations of Physical Activities 125

Moreover, the representation of victory, of what needs to be done in order

to win (as a team) was very significant in this network. Fundamentally, vol-
leyball was represented in this network as a competitive sport.
In Network A, the students’ representation seemed to be that in order to
progress and enjoy playing volleyball, you need to train. The main ideas we
found were progress, team spirit and enjoying the activity.
In Network B, the team dimension in volleyball was predominant and
students seemed to think that playing volleyball in teams required particular
tactics and qualities (mental and physical).
In Network C, the main representation that emerged was that of volleyball
as a team activity, where the goal was to attack and win (either using force
or dummy moves).
In our opinion, these three networks seemed to structure all the data that
was collected, in particular the relationships between the different variables
from the questions about team sports, the semantic differentiator and the
volleyball word association test. We shall now discuss which forms of sex and
gender variables had the greatest implications for these networks.

3.3 The influence of the additional “sex” and “gender” variables

on the networks revealed by the 0.70 threshold

For each chain, we calculated the influence of the additional variables: two
forms for sex (female and male) and four for gender (androgynous, non-
differentiated, feminine and masculine). Each of these additional variables had
a different influence on the formation of chains in the various networks. We
consider that the influence of an additional variable may be included in our
comments as an answer to our hypotheses, provided that its error rate is less
than 0.10 (a standard criterion for deciding whether a variable significantly
explains a particular phenomenon).

Influence of additional variables on Network A

When we looked at the chains that finished with: “play”, “feel good”, “control
yourself”, “play as a team” which were implied by the “I like team sports” vari-
able; we noticed that regardless of what the “previous” variable was (to return
it, to train, to relax, don’t get hurt, gentle), the female form of the additional
sex variable significantly influenced these chains (error rates: 0.0438, 0.0137,
0.042).
On the “a fun activity — feel good — control yourself — play as a team”
chain, it was again the female form of the sex variable which had a significant
influence, with an error rate of 0.0482.
The fact that the female form of the sex variable characterised all these
chains led us to hypothesise that females consider volleyball to be a PSA,
whose important aspects are team spirit, making progress and having fun.
126 C.-M. Chiocca and I. Verscheure

It therefore seemed that the representations in Network A were formed

around the idea that you first have to train and make progress before be-
ing able to enjoy volleyball. This would correspond more to the feminine
characteristics encountered when playing volleyball as described in the liter-
ature [6, 14].

Influence of additional variables on Network B

The female form of the sex variable characterised several chains.
• On the “a fun activity — a team sport — I like team sports” chain, the
female form of the sex variable had a significant influence, with an error
rate of: 0.0891.
• On the “a fun activity — I like volleyball” chain, the female form of the
sex variable had a significant influence, with an error rate of: 0.0764.
• On the “mental qualities — to play in continuity — I like team sports”
chain, the female form of the sex variable had a significant influence, with
an error rate of: 0.0863.
• On the “mental qualities — to play in continuity — body in motion” chain,
the female form of the sex variable had a significant influence, with an error
rate of: 0.0972.
However, two forms of the gender variable had a significant influence as ad-
ditional variables on two other chains:
• On the “mental qualities — a team sport — I like team sports” chain, the
feminine form of the gender variable had a significant influence, with an
error rate of: 0.0621.
• On the “physical qualities — I like volleyball — I like team sports” chain,
the masculine form of the gender variable had a significant influence, with
an error rate of: 0.0983.
This network’s representation, containing representations of volleyball as a
fun team activity where certain physical and mental qualities are needed,
seemed more comparable with the “female sex”. However, the “feminine gen-
der” had more influence where “mental qualities” were concerned; whereas
with “physical qualities” the “masculine gender” was predominant. We thus
suggest the idea of a network which is mainly mixed-sex in its formation or
representations.

Influence of additional variables on Network C

This network suggested that volleyball was represented as a team opposing
relationship, where the main interest came from playing matches and tack-
ling the opponent. Volleyball is all about team spirit, attacking and playing
matches.
The male form of the sex variable had a significant influence on several
chains in this network:
 Implicative networks of student’s representations of Physical Activities 127

• the “to win — to be ready to fight — I like team sports” chain, with an
error rate of: 0.0184.
• the “rough — to be ready to fight — I like team sports” chain, with an
error rate of: 0.0449.
• the “to win — attack — I like team sports” chain, with an error rate of:
0.041.
• the “to risk — I like team sports” chain, with an error rate of: 0.00572.
However, the androgynous form of the gender variable had a significant
influence on the following chains:
• “champion — to be ready to fight — I like team sports”, with an error rate
of: 0.0189.
• “champion — match — I like team sports” with an error rate of: 0.0246.
This network seemed more comparable with representations of an opposing
relationship in volleyball. The masculine form of the additional gender variable
had a significant influence on several chains, while the androgynous form was
predominant in one chain. It would rather be males, therefore, who would see
volleyball as a sport of continual opposition. The representation of adversity
(towards opponents as well as oneself) was revealed by this network in two
forms: an aggressive manner (“attack”, “rough”) and a cunning manner (“get
clever”, “to control yourself”, “to risk”). We therefore suggest that this network
was more “male” than the previous two.

4 Conclusion

In view of the results, we can say that the sex variable appeared more often
than the gender variable, as an additional variable with a significant influence
on these chains. Only the “feminine”, “androgynous” and “masculine” forms of
the gender variable each had a significant influence on a single chain (the non-
differentiated form of this variable did not predominate in any of the chains),
whereas both forms of the sex variable (“female” and “male”) played a greater
role.
Furthermore, Network B highlighted volleyball’s team dimension and the need
for implementing tactics and have particular mental and physical qualities.
This network, which could be summarized as: “playing volleyball as a team
requires tactics and mental qualities”, seemed closer to the “female” form; al-
though the “feminine gender” was implied when the answers stated that men-
tal qualities were required, while the “masculine gender” was implied when
physical qualities were concerned. The “female” form’s influence on this chain
should thus be moderated.
Network C was more synonymous with representations of volleyball as an op-
posing relationship. There was the significance of the matches, as well as the
desire to win, attack and play as a team. The additional variable having the
128 C.-M. Chiocca and I. Verscheure

greatest influence on this network was the “male” form, with the androgynous
gender also making a contribution to the “ready to fight” chain.
This level of analysis led us to believe that the students could have sexual rep-
resentations of volleyball, although sometimes one should make adjustments
according to gender.
In our other studies, not covered here, we drew parallels between classes that
were revealed using ascendant hierarchical classification (AHC) and networks
revealed using CHIC. We can therefore form student typologies according to
sex, that sometimes need to be adjusted to take gender into account.
While sex is a strong predictor of attitudes and dispositions to team sports
and volleyball, gender is not. We found that females and males prefer mascu-
line forms of team games, and that both see volleyball as relatively gender-
neutral. Both, females and males, appear to like volleyball, but for different
reasons. While the females enjoy the cooperative effort of keeping the game
going and so for continuity, the males show a strong preference for scoring
points and so for discontinuity.

References
1. J.-C. Abric. Pratiques sociales et représentations. PUF, Paris, 1994.
2. M. Bailleul. Mise en évidence de réseaux orientés de représentations dans deux
études concernant des enseignants stagiaires en iufm. In Actes des journées sur
la fouille dans les données par la méthode d’analyse statistique implicative, 2000.
3. S.L. Bem. The measurement of psychological androgyny. Journal of Consulting
and Clinical Psychology, vol. 42, n. 2:pp. 155–162, 1974.
4. D. Bouthier and B. David. Représentation et action: de la représentation initiale
à la représentation fonctionnelle des aps en eps. Méthodologie et didactique de
l’éducation physique et sportive, Ed. G. Bui-Xuan:pp. 233–249, 1989.
5. B. David. Rugby mixte en milieu scolaire. Revue Française de Pédagogie,
n. 110:pp. 51–61, 1995.
6. Davisse. Sport, école et société: la part des femmes. Paris Ed. Actio, pages
pp. 174–263, 1991.
7. Fontayne, Sarrazin, and Famose. The bem sex-role inventory: validation of a
short-version for french teenagers. European Review of Applied Psychology, 50,
n. 4:pp. 405–416, 2000.
8. R. Gras, S. Almouloud, M. Bailleul, and A. Larher. L’implication statistique.
Nouvelle méthode exploratoire de données. La Pensée Sauvage, 1996.
9. R. Gras and P. Kuntz. The implicative statistical analysis — its theoretical
foundations. Kluwer, 2007.
10. Jodelet. L’association verbale. in P. Fraisse et J. Piaget, Traité de psychologie
espérimentale, fasc. VIII:p/97–153, 1972.
11. R. Menahem. Le différenciateur sémantique, le modèle de mesure. L’année
psychologique, 68:pp. 451–465, 1968.
12. C.E. Osgood, G.J. Suci, and P.H. Tannenbaum. The mesurement of meanings.
Chicago, University of Illinois Press, 1957.
 Implicative networks of student’s representations of Physical Activities 129

13. M-L. Rouquette and P. Rateau. Introduction à l’étude des représentations so-
ciales. Grenoble, PUG, 1998.
14. Tanguy. Le volley: un exemple de mise en oeuvre didactique. Echanges et con-
troverses. n. 4:p. 7–20, 1992.
15. I. Verscheure, C. Amade-Escot, and C.-M. Chiocca. Représentations du volley-
ball scolaire et genre des élèves: pertinence de l’inventaire des rôles de sexe de
Bem? In RFP 154. 2006.
16. I. Verscheure and C. Amade-Escot. Gender difference in the learning of volley-
ball attack. In In Procedings AIESP Congress: Professionnal preparation and
social needs, 2002.

Appendix
The word association test was introduced in the following way: “What does
the word “volleyball” make you think about? Can you give me some other
words (between 5 and 7) that it brings to mind?”
The students gave us a total of 2386 words (an average of 4.7 words per
pupil); including 521 different words, with some being quoted very often. For
example: the word “net” was quoted 201 times, the word “ball” 198 times; the
word “smash” 175 times, the word “pass” 102 times and the word “team” 97
times. . .
To help process this information, we grouped the words together into sev-
eral categories. We began by grouping together words with the same root
(ball, volleyball), then words which seemed similar in terms of the research
questions, or in terms of their meaning. For example, we categorised the word
“team” together with other word groups including it, such as “good team at-
mosphere”, “team mates”, “team”, “solid team”, “be in a team”, “team game”,
“play as a team”, etc. We combined the “team” words with others relating
to the “collective” theme (e.g.: “learn to play together”, “good group spirit”,
“communal”, “colleagues”, “closeness”, “trust your partners”, “understanding
between players”, “mutual support”, “group”). This category was entitled “the
communal aspect of volleyball”. We continued in the same way for all 521
words, which were eventually grouped together into twenty categories.
To test the validity of grouping the words in this way, we called upon the
“judging method”. Two volleyball experts examined the words in our selected
categories, to let us know whether or not they agreed with the categories the
words had been put into. There was agreement on more than 80 per cent of
the words, so we retained the classification and, after discussion, revised the
word categorisation until a consensus was reached.

Osgood’s semantic differentiator was introduced in the following way:

Here are a series of terms evoking volleyball (in a broad sense). According
to which term at either end of the scale means the most to you, circle one
figure only for each line.
130 C.-M. Chiocca and I. Verscheure

Categories Number of occurences

A fun activity 96
Team spirit aspect 287
Attack 274
Cooperation 214
Difficult 21
Sexual aspect 25
Equipment/material (the ball in particular) 258
Movement/energy 44
Fear/pain 55
Mental qualities 44
Physical qualities 78
Opposing relationship 103
Reference to beach games 87
Rules/limits (the net in particular) 306
Negative feelings 38
Positive feelings 49
Tactics 99
Technique 210
Upwards 42
Not initialised 24

Return 3 2 1 0 1 2 3 Attack
Make progress 3 2 1 0 1 2 3 Become athletic
Score a point 3 2 1 0 1 2 3 Play as a team
Prolong the exchange 3 2 1 0 1 2 3 Interrupt the exchange
Become strong 3 2 1 0 1 2 3 Get clever
Be the best 3 2 1 0 1 2 3 Learn to control yourself
Watch the ball 3 2 1 0 1 2 3 Watch the opponent
Rough 3 2 1 0 1 2 3 Gentle
to be ready to fight 3 2 1 0 1 2 3 Don’t get hurt
Play 3 2 1 0 1 2 3 to win
Be a champion 3 2 1 0 1 2 3 Feel good
Static 3 2 1 0 1 2 3 Mobile
Make progress 3 2 1 0 1 2 3 Relax
to train 3 2 1 0 1 2 3 Match
Precision 3 2 1 0 1 2 3 Force
Be good enough 3 2 1 0 1 2 3 to risk
Rupture 3 2 1 0 1 2 3 to play in continuity

Table 1. Classification of words into twenty categories

A comparison between the hierarchical
clustering of variables, implicative statistical
analysis and confirmatory factor analysis

Iliada Elia and Athanasios Gagatsis

Department of Education, University of Cyprus

P.O. Box 20537, 1678 Nicosia Cyprus
{ilada, gagatsis}@ucy.ac.cy

Summary. This study aims to gain insight about the distinct features and advan-
tages of three statistical methods, namely the hierarchical clustering of variables,
the implicative method and the Confirmatory Factor Analysis, by comparing the
outcomes of their application in exploring the understanding of function. The inves-
tigation concentrates on the structure of students’ abilities to carry out conversions
of functions from one mode of representation to others. Data were obtained from
587 students in grades 9 and 11. Using Confirmatory Factor Analysis, a model, that
provides information about the significant role of the initial representations of con-
versions in students’ processes, is developed and validated. Using the hierarchical
clustering and implicative analysis, evidence is provided to students’ compartmen-
talized thinking among representations. These findings remain stable across grades.
The outcomes of the three methods were found to coincide and to complement each
other.

Key words: implicative analysis, hierarchical clustering, CHIC, Confirmatory Fac-

tor Analysis, function, representation.

1 Introduction
Nowadays the centrality of representations in teaching, learning and doing
mathematics seems to become widely acknowledged. A basic reason for this
emphasis is that mathematical concepts are accessible only through their semi-
otic representations [1]. Kaput suggests that representations are “integrated”
with mathematics. In certain cases, representations, such as graphs, are so
closely connected with a mathematical concept, such as function, that it is
difficult for the concept to be understood and acquired without the use of the
corresponding representation [2].
A given representation, however, cannot describe thoroughly a mathemat-
ical concept, since it highlights only a part of its aspects [3]. This justifies

I. Elia and A. Gagatsis: A comparison between the hierarchical clustering of variables,

implicative statistical analysis and confirmatory factor analysis, Studies in Computational
Intelligence (SCI) 127, 131–162 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
132 I. Elia and A. Gagatsis

the strong support in the mathematics education community that students

can grasp the meaning of mathematical concepts by experiencing multiple
mathematical representations [3–6].
The concept of function is central to mathematics and its applications.
It emerges from the general inclination of humans to connect two quantities,
which is as ancient as mathematics itself. There is a large body of literature
on the understanding of functions that focuses mainly on the role of different
representations. For example, a number of studies have shown that students
tend to have difficulties in transferring information related to functions gained
in one representational context to another [1, 3].
Recent studies have used different methods of analysis to investigate stu-
dents’ abilities and their structure in using various representations of function.
A few studies have employed the hierarchical clustering of variables [7,8], while
other studies have used the hierarchical classification in combination with the
implicative method [3, 9, 10]. There are also a number of studies that have
attained their outcomes by using Confirmatory Factor Analysis (CFA) [11].
An important question that arises is whether each of these studies would have
resulted in similar or congruent findings, if statistical methods other than the
one applied, were conducted on its data. Another crucial issue concerns the
aspects of a study that each statistical analysis serves better and helps more
efficiently to make sense of. In an attempt to tackle these questions, the in-
tent of this study, which focuses on students’ abilities in transferring functions
from one representation to another, is to apply all of the three aforementioned
statistical methods of analysis on the same sample data and compare their
outcomes.

2 Theoretical considerations: The role of representations

on the understanding of functions
Students experience a wide range of representations from their early childhood
years onward. A main reason for this is that most mathematics textbooks
today make use of a variety of representations more extensively than ever
before in order to promote understanding.
The use of multiple representations has been strongly connected with the
complex process of learning in mathematics, and more particularly, with the
seeking of students’ better understanding of important mathematical con-
cepts [12, 13], such as function. Given that a representation cannot describe
fully a mathematical construct and that each representation has different
advantages, using various representations for the same mathematical situa-
tion is at the core of mathematical understanding [1]. Ainsworth, Bibby and
Wood [14] suggest that the use of multiple representations can help students
develop different ideas and processes, constrain meanings and promote deeper
understanding. By combining representations students are no longer limited
by the strengths and weaknesses of one particular representation. Kaput [15]
 A comparison between the hierarchical clustering of variables 133

claims that the use of more than one representation or notation system helps
students to obtain a better picture of a mathematical concept.
The ability to identify and represent the same concept through different
representations is considered as a prerequisite for the understanding of the
particular concept [1, 16]. Besides recognizing the same concept in multiple
systems of representation, the ability to manipulate the concept with flexibility
within these representations as well as the ability to “translate” the concept
from one system of representation to another are necessary for the mastering
of the concept [5] and allow students to see rich relationships [16].
Duval [1, 17] maintains that mathematical activity can be analyzed based
on two types of transformations of semiotic representations, i.e. treatments
and conversions. Treatments are transformations of representations, which
take place within the same register that they have been formed in. Conversions
are transformations of representations that involve the change of the register
in which the totality or a part of the meaning of the initial representation is
conserved, without changing the objects being denoted.
Some researchers interpret students’ errors as either a product of a defi-
cient handling of representations or a lack of coordination between represen-
tations [13, 18]. The standard representational forms of some mathematical
concepts, such as the concept of function, are not enough for students to con-
struct the whole meaning and grasp the whole range of their applications.
Mathematics instructors, at the secondary level, traditionally have focused
their teaching on the use of the algebraic representation of functions [19].
Sfard [20] showed that students were unable to bridge the algebraic and graph-
ical representations of functions, while Markovits, Eylon and Bruckheimer [21]
observed that the translation from graphical to algebraic form was more dif-
ficult than the reverse. Sierpinska [4] maintains that students have difficulties
in making the connection between different representations of functions, in
interpreting graphs and manipulating symbols related to functions. Gagatsis
and Christou [11] developed a model involving some critical paths relating the
conversions from one type of representation to another. These paths indicated
that the conversion from one representation to another is not a straightforward
task. For example, students’ ability to translate a function from its graphical
to the algebraic form was the result of students’ understanding of three other
conversions: (a) the conversion of a function from graphic to verbal form,
(b) the conversion from verbal to graphic function, and (c) the conversion
from verbal to algebraic form of a function. A possible reason for this kind
of behaviour is that most instructional practices limit the representation of
functions to the translation of the algebraic form to the graphic form and not
the reverse. Furthermore, Aspinwall, Shaw and Presmeg [22] suggested that
in some cases the visual representations create cognitive difficulties that limit
students’ ability to translate between graphical and algebraic representations.
Lack of competence in coordinating multiple representations of the same
concept can be seen as an indication of the existence of compartmentaliza-
tion, which may result in inconsistencies and delays in mathematics learning
134 I. Elia and A. Gagatsis

at school. This particular phenomenon reveals a cognitive difficulty that arises

from the need to accomplish flexible and competent translation back and forth
between different modes of mathematical representations [1]. Elia et al. [10] ex-
amined whether pupils of grade 9 (14 year olds) accomplished the conversions
among different modes of representation of functions (i.e. graphic, symbolic,
verbal). They found that different types of conversion among representations
of the same mathematical content were approached in a completely distinct
way. These results provided a strong indication of pupils’ compartmentalized
thinking and use of the various representations of functions and therefore their
deficiencies in the understanding of the concept.

3 Aim and research questions

The aim of the study is to combine and compare the outcomes of CFA, hierar-
chical clustering of variables and implicative method on the same sample data
concerning students’ abilities in carrying out conversions among different rep-
resentations of functions. A main concern is to gain insight about the distinct
features, advantages and limitations of each of the three statistical methods
in a central topic of mathematics education, namely the understanding of the
concept of function, and to examine whether they coincide or even comple-
ment each other through students’ observed performance on this particular
subject.
In the light of the above, the following research questions are formulated:
• Which are the common features of the outcomes of the three statistical
methods? To what extent there is consistency between the results derived
from these processes?

• For which aspects of the study is the application of each statistical method
more appropriate and open to complementary use?

4 Method

4.1 Participants, instrument and variables

The sample of the study consisted of 587 students of grades 9 and 11 in

Greece. Specifically, 183 of the students were of grade 9 (14 years of age) and
404 of them were of grade 11 (16 years of age). Two tests were constructed and
administered to the participants. The first test (A) consisted of six tasks in
which students were given the graphic representation of an algebraic relation
and were asked to translate it to the verbal and symbolic form respectively.
The second test (B) consisted of six tasks (involving the same algebraic re-
lations with test A) in which students were asked to translate each relation
 A comparison between the hierarchical clustering of variables 135

from its verbal representation to the graphical and to the symbolic represen-
tation respectively. Students had to carry out 12 conversions in each test, that
is, 24 conversions in total. For each type of conversion the following types of
algebraic relations were examined: y < 0, xy > 0, y > x, y = −x, y = 3/2,
y = x − 2 based on a relevant research of Raymond Duval [23]. The former
three tasks correspond to inequalities and thus regions of points, while the
latter three tasks correspond to functions.
Each test included an example of an algebraic relation in a graphic, verbal
and symbolic form to help students understand what they were asked to do.
The example is illustrated in Table 1.

Graphic Verbal Symbolic

representation representation representation

It represents the region of x > 0

the points having positive
abscissa.

Table 1. An example of the tasks included in the test

Correct responses to the tasks were assigned a score of 1, while incorrect

answers were given a score of 0. The variables used for the analyses of the
data corresponded to students’ responses to the tasks and were symbolized as
follow: V11a, V12a, V21a, V22a, V31a, V32a, V41a, V42a, V51a, V52a, V61a,
V62a, V11b, V12b, V21b, V22b, V31b, V32b, V41b, V42b, V51b, V52b, V61b
and V62b. The symbolism used for the variables of the data is explained below:
(a) “a” stands for Test A, and “b” stands for Test B
(b) The first number after “v” stands for the number of the task in the test
i.e. 1: y < 0, 2: xy > 0, 3: y > x, 4: y = −x, 5: y = 3/2, 6: y = x − 2
(c) The second number stands for the type of conversion for each test,
i.e. for Test A, 1: graphic to verbal representation, 2: graphic to symbolic
representation; for Test B, 1: verbal to graphic representation, 2: verbal to
symbolic representation.

4.2 Data analysis

This section is distinguished into two parts. The first part involves a brief
overview of the rationale, the components and basic concepts of structural
136 I. Elia and A. Gagatsis

equation modeling and CFA, while the second part concentrates on the under-
lying principles, elements and structure of the implicative statistical method
and the hierarchical classification of variables.

Structural Equation Modeling and CFA

“Structural equation modeling (SEM) is a statistical methodology that takes a

hypothesis testing (i.e. confirmatory) approach to the multivariate analysis of
a structural theory bearing on some phenomenon” [24]. This theory concerns
“causal” relations among multiple variables [25]. These relations are repre-
sented by structural, namely regression equations, which can be modeled in a
pictorial way to allow a better conceptualization of the involved theory.
SEM differs from the more traditional multivariate statistical techniques
in at least three dimensions: First, with the use of SEM the analysis of the
data is approached in a confirmatory manner rather than in an exploratory
way, making hypothesis testing more accessible and easier, compared with
other multivariate procedures. Second, whereas SEM gives the estimates of
measurement errors, the “conventional” multivariate methods cannot assess or
correct for these parameters. Third, SEM involves not only observed but also
latent (unobserved) variables, whereas the older techniques incorporate only
observed measurements.
Latent and observed variables are two of the most basic concepts of SEM.
Latent variables (i.e. factors) represent theoretical or abstract constructs that
can not be observed or measured directly and are rather assumed to lie behind
certain observed measures. Examples of latent variables in this study are the
abilities of students to convert functions from one mode of representation
to another. The measurement of latent variables is obtained indirectly by
associating it with other variable/s that is/are observable. The latent variable
is based on some behaviour supposed to represent it. This behaviour refers to
scores on a particular instrument, like the test of this study, and in turn these
measured scores are called observed variables.
Factor analysis is a well known statistical technique for examining asso-
ciations between observed and latent variables. The covariation among a set
of observed variables is investigated to get information on their underlying
latent factors. Of primary interest is the strength of the regression paths from
the factors to the observed variables. Exploratory Factor Analysis (EFA) and
Confirmatory Factor Analysis (CFA) are two basic types of factor analysis.
EFA is employed to determine how the observed variables are connected to
their underlying constructs in situations where these links are unknown. By
contrast, CFA, which is the type of analysis employed here, is used in situa-
tions where the researcher aims to test statistically whether a hypothesized
linkage pattern between the observed variables and their underlying factors
exists. This a priori hypothesis draws on knowledge of related theory and past
empirical work in the area of the study. The basic steps that a researcher fol-
lows for carrying out CFA are described below: The model is specified based on
 A comparison between the hierarchical clustering of variables 137

knowledge of relevant theory and previous empirical research. Using a model-

fitting program, such as EQS, the model is analyzed so that the estimates of
the model’s parameters with the data are derived. Then the tenability of the
model is tested based on data that involve all the observed variables of the
model. In other words, what is tested is how well the observed data fit the a
priori structure. If the hypothesized model is not consistent with the data the
model is respecified and the fit of the revised model with the same data is
evaluated [24, 26].
The number of levels that the latent factors are away from the observed
variables determines whether a factor model is called a first-order, a second-
order or a higher order model. Correspondingly, factors one level removed
from the observed variables are labeled first-order factors while higher-order
factors which are hypothesized to account for the variance and co-variance
related to the first-order factors are termed second-order factors. A second or
a higher order factor does not have its own set of measured variables. In this
study a second-order and a third-order model will be considered. A structural
equation model involves two basic types of components: the variables and the
processes or relations among the variables. A schematic representation of a
model, which is termed path diagram, provides a visual interpretation of the
relations that are hypothesized to hold among the variables under study. The
basic notation and schematic presentation of the variables and the relations
that are used in path diagrams are described below.
The observed or measured variables, which constitute the actual data of
the study, are often designated as Vs and are shown in rectangles. The unmea-
sured variables, which are hypothetical and represent the structural organiza-
tion of the phenomenon under study, are of three types: a) the latent factor
which is designated as F and represented in the path diagram in ellipses or cir-
cles; b) a residual associated with the measurement of each observed variable
which is referred to as error and designated as E; and c) a residual or the er-
ror associated with the prediction of each factor, which is termed disturbance
and designated as D. Residual terms indicate the imperfect measurement of
the observed variables and the imperfect prediction of the unobserved factor.
Although both kinds of residuals represent errors, the former is termed er-
ror, and the latter is referred to as disturbance, to distinguish the error in
measurement from error in prediction. For example, in the model of Figure 1,
“V112a”-“V612a” and “V112b-V612b” represent the measured variables, “Gvs”,
“Vgs” and “Gvs” stand for the latent factors, “E1”–“E12” correspond to the
residuals of the observed variables and “D1”–“D12” refer to the residuals of
factors.
One type of the relations involved in a model is the structural regression
coefficients indicating the impact of one variable on another. They are repre-
sented by one-way arrows. For example, the unidirectional arrows leading from
the Factor “Gvs” (Figure 1) to the six observed variables (V112a-V612a) in-
dicate that the scores on the latter variables are “caused” by the factor “Gvs”.
These relations are called “factor loadings”. Similarly, unidirectional arrows
138 I. Elia and A. Gagatsis

from one factor to another imply that a factor causes or predicts another fac-
tor, e.g., in Figure 1 the arrows starting from “AbGV” and pointing toward
“Gvs” and “Vgs” imply that “AbGV” predicts “Gvs” and “Vgs”. A second type
of processes used in a model is the impact of the errors in the measurement of
the observed variables and in the prediction of the latent factors. The impact
of random measurement errors on the observed variables and errors in the pre-
diction of factors are represented as one-way arrows pointing from Es (e.g.,
E1-E12) and Ds (e.g., D1, D2) respectively, to the corresponding variables,
as shown in Figure 1. A third type of processes involved in a model are the
covariances or correlations between pairs of variables, which are represented
as curved or (sometimes) straight two-way arrows (e.g., E1-E2 or E7-E8) as
illustrated in Figure 1.
Bentler’s [27] EQS program was used for testing the CFA models in this
study. The estimation method that was used in EQS was maximum likelihood
solution. The tenability of a model can be determined by using the following
measures of goodness-of-fit: X 2 , CFI (Comparative Fit Index) and RMSEA
(Root Mean Square Error of Approximation) [28]. The following values of the
three indices are needed to support model fit: The observed values for X 2 /df
should be less than 3, the values for CFI should be higher than .9 and the
RMSEA values should be lower than .06.

Implicative Statistical Analysis and Hierarchical Clustering

of Variables

For the analysis of the collected data, the hierarchical clustering of variables
and Gras’s implicative statistical method have been also conducted using a
computer software called C.H.I.C. (Classification Hierarchique, Implicative et
Cohesitive), Version 3.5. [29]. These methods of analysis determine the hier-
archical similarity connections and the implicative relations of the variables
respectively [30, 31]. For this study’s needs, similarity and implicative dia-
grams have been produced from the application of the analyses on the whole
sample and on each age group of students. The implications of the analyses
were based on the classical theory.
The hierarchical clustering of variables [32] is a classification method which
aims to identify in a set V of variables, sections of V, less and less subtle,
established in an ascending manner. These sections are represented in a hier-
archically constructed diagram using a similarity statistical criterion among
the variables. The similarity stems from the intersection of the set V of vari-
ables with a set E of subjects (or objects). This kind of analysis allows the
researcher to study and interpret clusters of variables in terms of typology
and decreasing resemblance. The clusters are established in particular levels
of the diagram and can be compared with others. This aggregation may be
indebted to the conceptual character of every group of variables.
The construction of the hierarchical similarity diagram is based on the
following process: Two of the variables that are most similar to each other
 A comparison between the hierarchical clustering of variables 139

with respect to the similarity indices of the method are joined together in a
group at the highest (first) similarity level. Next, this group may be linked
with one variable in a lower similarity level or two other variables that are
combined together and establish another group at a lower level, etc. This
grouping process goes on until the similarity or the cohesion between the
variables or the groups of variables gets very weak. In this study the similarity
diagrams allow for the arrangement of the variables, which correspond to
students’ responses in the tasks of the tests, into groups according to their
homogeneity.
The implicative statistical analysis aims at giving a statistical meaning to
expressions like: “if we observe the variable a in a subject, then in general
we observe the variable b in the same subject” [30, 33]. Thus the underlying
principle of the implicative analysis is based on the quasi-implication: “if a is
true then b is more or less true”. An implicative diagram represents graphically
the network of the quasi-implicative relations among the variables of the set V.
In this study the implicative diagrams contain implicative relations, which
indicate whether success to a specific task implies success to another task
related to the former one.
It should be noted that the present paper is related to the ones of Elia et
al. [10] and Gagatsis and Christou [11], whose basic findings are included in
the theoretical section (2).

5 Results

5.1 The outcomes of CFA

Before carrying out CFA, we examined the hypothesis that the data of our
sample come from a normal population. The values of skewness and kurtosis,
each divided by the corresponding standard errors, for the whole sample (0,6
and -3,2) and for each age group (Grade 9: 1,9 and -1,9; Grade 11: 0,0 and
-2,1) indicated that the data were normally distributed.
Next, a series of models were tested and compared. Specifically, the first
model was a third-order CFA model which was designed on the basis of the
results of the study by Elia et al. [10]. It involved one third-order factor which
was hypothesized as accounting for all variance and covariance related to the
second-order factors. The second-order factors represented students’ abilities
to carry out conversions of algebraic relations with the graphic (Test A) and
the verbal mode (Test B) respectively as the source representation. Each of
the second-order factors were assumed to explain the variance and covariance
related to three first-order factors measured by the observed variables cor-
responding to the 12 conversions of Test A and the 12 conversions of Test
B respectively. The former three first-order factors were distinguished with
respect to the conceptual characteristics of the tasks, that is, whether they
involved a function or not or the kind of the function involved. The latter
140 I. Elia and A. Gagatsis

three first-order factors were differentiated with reference to their conceptual

characteristics, but also to their type of conversion and more specifically their
target representation (graphic or symbolic). The fit of this model was poor
[X 2 (244) = 957.889; CFI = .831; RMSEA = .071, 90% confidence interval for
RMSEA = 0.066–0.075], indicating that the particular structure was not ap-
propriate to describe the structure of the abilities in performing conversions
among the different modes of representation of the algebraic relations included
in the two tests. The structure of this model was based on a combination of
the results of the two age groups given by Elia et al. [10]. Consequently, some
relationships in the model that applied for grade 9 did not apply for grade
11 and vice versa. This could be an explanation for the inconsistency of this
model with the data of the whole sample.
Nevertheless, a main concern of this study was to validate a CFA model
that could capture the structural organization underlying the processes of the
students of both age groups in the conversions among different representations
of functions. Therefore, a critical commonality that was observed between the
two age groups in Elia et al.’s [10] study was the inconsistency in students’
performance when dealing with conversions of a different source representa-
tion. This strong tendency among the students of both age groups, as well as,
Duval’s findings suggesting that a major difficulty in mathematics learning is
the compartmentalization among registers of representations, formed a basis
for the second model that we have examined [1]. The second model (Figure 1)
involves two first-order factors and one second-order factor on which the first
order-factors are regressed. The first-order factors stand for the two types of
conversions of the six algebraic relations involved in the two tests with respect
to their initial (source) representation, i.e. conversions from a graphic form
to a verbal and to a symbolic form (Gvs) and conversions from a verbal form
to a graphic and to a symbolic form (Vgs). Each factor was measured by six
observed variables, which stand for the conversion tasks of the six mathemat-
ical relations having a common initial representation. In Figure 5.1 we have
to note that:
1. “AbGV” stands for the ability to carry out conversions of functions and
other algebraic relations having the graphic or the verbal form as the
source representation.
2. “Gvs” stands for the ability to carry out conversions of functions and other
algebraic relations from the Graphic form to the verbal and the symbolic
form.
3. “Vgs” stands for the ability to carry out conversions of functions and other
algebraic relations from the Verbal form to the graphic and the symbolic
form.
4. E1–12 stand for the errors of the variables.
5. D1 and D2 stand for the errors of the latent factors.
 A comparison between the hierarchical clustering of variables 141

Fig. 1. The elaborated model for the conversions among different modes of repre-
sentation of algebraic relations, with factor loadings for students of the whole sample
and of grades 9 and 11, separately.

6. The first, second and third coefficient of each parameter stand for the
application of the model on the performance of the students of the whole
sample, of grades 9 and 11, respectively.
It is noteworthy that the number of the observed variables is half as the num-
ber of the observed variables of the first model. In the tests the conversion
items of each algebraic relationship formed one task, having the same starting
representation, and students normally treated the two conversion items of the
same relationship “simultaneously”. Therefore, we considered that it would be
more meaningful to integrate the variables that corresponded to the conver-
sions of the same mathematical relation and the same initial representation,
142 I. Elia and A. Gagatsis

despite their difference in the target representation. We combined each pair of

variables (e.g., V11a and V12a) involving the same relation in the same test
into one variable (e.g., V112a).
The second-order factor represents the general ability to perform conver-
sions of algebraic relations starting from a graphic or a verbal form of represen-
tation (AbGV). The fit of this model was good [X 2 (50) = 165.730; CFI = .943;
RMSEA = .063, 90% confidence interval for RMSEA = 0.052–0.073], verifying
that the type of the conversion and more specifically the type of the starting
representation of a conversion does have an effect on the process of carry-
ing out conversions of functions and other algebraic relations among different
representations.
As written above, combining the variables, the degrees of freedom of the
models are decreased from 250 to 50. This is due to the fact that the number
of data points, i.e. variances, covariances of the observed variables (how much
information we have with respect to our data), is significantly decreased. In
particular, while the data points of the model involving 24 observed variables
were 300 (24 × 25/2), the corresponding elements of the model involving half
of the variables were 78 (12 × 13/2). The number of estimable parameters in
the former model was 50, while the number of parameters to be estimated
in the latter model was 28. Given that the degrees of freedom stand for the
difference between the data points and the structural parameters, the former
model has 250 degrees of freedom, while the latter one has only 50.
Attention is drawn to the fact that the relation of the first-order factor
standing for the conversion having the graphic mode as the starting repre-
sentation to the second-order factor, corresponding to the general ability in
representational conversions of functions, is considerably stronger than the re-
lation of the other first-order factor. Thus, the conversions involving the verbal
mode as the starting representation seem to have considerable autonomy from
the conversions involving the graphic form as the source representation.
To test for possible differences between the two age groups in the structure
described above, confirmatory factor analysis was applied to test the second-
order model separately on each ge group. The model was found to be con-
sistent with the data of both groups [grade 9: X 2 (50) = 108.354; CFI = .911;
RMSEA = .082, 90% confidence interval for RMSEA = 0.061–0.102, grade 11:
X 2 (50) = 141.846; CFI = .926; RMSEA = .068, 90% confidence interval for
RMSEA = 0.054–0.080].
The regression coefficients of the scores in the conversion tasks of the al-
gebraic relations standing for regions of points (V112a, b: y < 0, V212a, b:
xy > 0) onto the corresponding first-order factors representing students’ abil-
ities in the conversions having as a source representation the graphic or the
verbal form were lower than the regression coefficients of the other variables.
Thus, the conversion of these algebraic relations seems to have considerable
autonomy from the conversions of relations involving functions. Moreover,
students’ abilities to carry out the conversions of these inequalities are sig-
nificantly correlated to each other in each test. The regression coefficient of
 A comparison between the hierarchical clustering of variables 143

the conversion of the constant function (V512a: y = 3/2) with the graphic
form as the source representation was also lower compared to the respective
coefficients of the variables standing for the other functions in Test A. This is
an interesting finding to be discussed later in combination with the results of
the other statistical methods.

5.2 The outcomes of the hierarchical clustering of variables

and the implicative method of analysis

Each observed variable in the CFA model (Figure 1) represented students’

“unified” score at the conversions of each algebraic relation from the graphic
representation to the verbal and the symbolic representation (Test A) or from
the verbal representation to the graphic and the symbolic representation (Test
B). To make the comparison between the three statistical techniques under
study feasible, we considered that it would be meaningful to include these
variables also in the implicative analysis and the hierarchical classification.
Before elaborating on the results of the hierarchical clustering of variables
and the implicative method, we will explicate our predictions concerned with
the structure of the similarity and implicative diagrams involving the variables
of the study. First, we expect that similarity and implicative relationships will
be primarily established among the variables corresponding to the conversions
of the same starting representation, namely the graphic and the verbal repre-
sentations. This hypothesis is based on well documented findings suggesting
the fragmentary way of students’ thinking when dealing with different types
of representations [1,6]. A second prediction is that close relationships will be
formed among variables standing for the conversion tasks of functions because
of their common perceptual and conceptual features. Third, we expect that
success on the conversion of functions would entail success on the conversions
of inequalities in the implicative diagrams. We consider the understanding of
inequalities (graphically, regions of points) and principally of the inequality
y < 0 and other similar ones as prerequisite for the understanding and use of
more complex relations and functions such as y = −x or y = x − 2.
The symbolism used for the variables of figure 2 (and the figures that
follow) is the following:
1. “a” stands for Test A, and “b” stands for Test B.
2. The first number after “v” stands for the number of the task in the test.
i.e. 1: y < 0, 2: xy > 0, 3: y > x, 4: y = −x, 5: y = 3/2, 6: y = x − 2
3. The second number stands for the source representation of the conversion
corresponding to each test, i.e. for Test A, 1: graphic representation; for
Test B: verbal representation.
Figure 2 illustrates the similarity relations among the “condensed” vari-
ables corresponding to grade 9 and 11 students’ responses to the tasks of the
two tests. Two distinct clusters of variables are established in the hierarchical
similarity diagram. The first cluster involves students’ responses to the tasks
144 I. Elia and A. Gagatsis

Fig. 2. The hierarchical similarity diagram among the responses of students of

grades 9 and 11 to the conversion tasks of Test A and Test B (C.H.I.C. 3.5)

of Test A, while the second cluster is comprised of students’ responses to the

tasks of Test B. This suggests that students dealt with the conversions of the
six algebraic relations which had the graphic mode as the initial representation
consistently. Students exhibited consistency, in a lower level though, also in
the conversions of the corresponding algebraic relations which had the verbal
form as the initial representation. These findings verify the first prediction
concerned with the variables’ relationships, suggesting the establishment of
close connections among students’ responses to the conversions involving the
same starting representation. The weak similarity relation between the two
clusters suggests that students approached conversions of a different source
representation in a distinct way, despite involving the same algebraic relations.
This behaviour may be a consequence of students’ compartmentalized way of
thinking in different modes of representation.
The similarity relations within each cluster of variables are also of great in-
terest since they can be seen as indications of students’ way of understanding
of the particular algebraic relations and of carrying out the particular types of
conversion. Each cluster involves one similarity group, namely Groups 1a and
2b, in which the variables of the third task (v312a or v312b), the fourth task
(v412a or v412b), the fifth task (v512a or v512b) and the sixth task (v612a
or v612b) are linked together. On one hand, the similarity linkage among the
variables referring to tasks 4, 5 and 6 which involve function provides evidence
to the second prediction of our analysis, stating that close relationships will
be established among variables standing for the conversion tasks of functions.
On the other hand, the similarity connection of these variables with the vari-
able referring to the conversion task 3 of the inequality y > x extends our
 A comparison between the hierarchical clustering of variables 145

expectation. It is suggested that students carried out the conversions of

functions, that is, tasks 4–6, as well as the algebraic relation which involved
a function, that is, task 3, using similar processes. The conversions of the
graphic or the verbal form of y < 0 (v112a or v112b) and of xy > 0 (v212a
or v212b) were carried out differently from the conversions of the other rela-
tions probably due to their distinct properties as they did not entail functions.
Whereas students tackled the conversions of the relations y < 0 (v112b) and
xy > 0 (v212b) having the verbal form as a source representation similarly to
each other thus forming a similarity pair (Group 2a), this was not the case in
the conversions of the respective relations (v112a and v212a) with the graphic
form as the initial representation. This finding provides further support to the
assertion that the graphic form and the verbal form of the same mathematical
content stimulated the use of distinct conversion processes by the students.
Figure 3 shows the implicative relations among the “condensed” variables
corresponding to ninth and eleventh graders’ responses to the tasks of the two
tests. The diagram involves a dual implicative chain which indicates the hier-
archical ordering of the conversion tasks with respect to their level of difficulty
on the basis of students’ performance. One branch of the chain, namely Branch
A, involves mainly the responses to the conversions of the algebraic relations
with the graphic representation as the initial one (Test A). The other branch,
termed Branch B, is comprised mainly of the responses to the conversions of
the algebraic relations with the verbal representation as the initial one (Test
B). The establishment of these implicative branches gives further support to
the first prediction for the analyses stating that implicative relationships will
be primarily formed among the variables corresponding to the conversions of
the same starting representation, namely the graphic and the verbal form.
Both branches stem from the same variable which stands for the response to
the conversion of the graphic representation of the function y = x − 2 (v612a).
This is the most complex task of both tests and students who provided a cor-
rect solution at it, succeeded at all of the other conversion tasks of both tests.
Students’ great difficulty to the sixth conversion task of each test is verified
also by their lowest success rates on the particular task in both grades as
illustrated in Tables 5 and 6 of the Appendix. These tables present the means
and frequency of occurrences of the two age groups, respectively, to all of the
tasks of the two tests.
Considering “Branch B”, students who accomplished the most difficult task
of both tests succeeded at the corresponding task in Test B, meaning the con-
version of the verbal representation of the same algebraic relation (v612b).
Carrying out the latter task implied success at the task of the same type
of conversion involving the function y = −x (v412b), which in turn entailed
correct performance in the conversions of the function y = −3/2 (v512b).
These implicative relationships give further support to the second prediction
for the analyses suggesting the formation of implicative relationships among
variables standing for the conversion tasks of functions. The conversions of the
constant function from a verbal to a symbolic or to a graphic representation
146 I. Elia and A. Gagatsis

Fig. 3. The implicative diagram among the responses of students of grades 9 and
11 to the conversion tasks of Test A and Test B (C.H.I.C. 3.5)

seemed to be easier than the corresponding conversions of any other function

involving the variable of x. The conversions of functions were more difficult,
though, than the conversions of inequalities standing for regions of points in
Test B (see Appendix). Moreover, success at the conversions of the verbal rep-
resentation of functions implied success at the conversions of the same source
representation of algebraic inequalities. Success at the conversions of the con-
stant function, entailed success at the conversions of the algebraic relation
y > x (v312b), which sequentially implied success at the corresponding con-
versions of the relation xy > 0 (v212b). Students who performed correctly at
the latter task, were able to carry out the simplest conversion task of Test B
which involved the relation y < 0 (v112b). Therefore, our third prediction that
success on the conversions of functions would entail success on the conversions
of inequalities in the implicative diagrams was also verified.
Considering “Branch A”, carrying out the most difficult task (v612a) en-
tailed success at the conversions of the function y = −x from a graphic
to a verbal and to a symbolic representation (v412a). Success at the latter
task implied success at the easier task incorporating the algebraic relation
 A comparison between the hierarchical clustering of variables 147

y > x (v312a). Consecutively, students who carried out the conversions of the
graphic representation of y > x were successful at the corresponding conver-
sions of the constant function y = 3/2 (v512a). Success at the latter conversion
task implied success at the conversions of the relation xy > 0 (v212a) corre-
sponding to a region of points, which in turn entailed success at the conversion
of the graphic form of the relation, y < 0 (v112a), representing a region of
points as well. On one hand, these results indicate that in general the hier-
archical ordering of the tasks based on students’ performance to Test A is
congruent with Test B, providing further evidence to the two latter predic-
tions which suggest the establishment of close implicative relationships among
the variables of the conversions of functions and of the implications of success
at these conversions to success at the conversions of regions of points. On the
other hand, unlike Branch B, the variable of the conversion of y > x intervenes
among the variables of the conversions of functions in Branch A, which is not
completely in line with these predictions. The conversion of the graphic repre-
sentation of the relation y > x, which despite being an inequality, it actually
involves a function, was more complex than the corresponding conversion of
the constant function and success at the task of the former relation implied
success at the task of the latter relation.
It is worth noting that besides the implicative relation between the vari-
ables referring to the two conversion tasks that involved the same algebraic
relation y = x − 2 (v612a-v612b), there is another implicative connection link-
ing variables of the two tests. In particular, success at the conversions of the
relation xy > 0 having the graphic mode as the initial representation (Test
A: v212a) implied success at the conversion of the same relation with the
verbal form as the initial representation (Test B: v212b). On one hand, the
fact that only two implicative relations are formed between the variables of
the two tests suggests that students’ success at most of the tasks of Test A
was independent from their success at the tasks of Test B. Thus, support is
provided to the almost compartmentalized ways by which the students ap-
proached the conversions of a different source representation despite involving
the same mathematical content. On the other hand, the fact that in both
relations success at a task of Test A entailed success at a task of Test B pro-
vides evidence to the more difficult character of the conversion starting with
a graph relatively to a conversion starting with a verbal description.
The next figures illustrate the results of the hierarchical clustering of vari-
ables and the implicative method for the students of Grade 9 and 11 sepa-
rately. These results are generally congruent with the outcomes referring to
the whole sample elaborated above, and therefore they are in line with the
predictions concerned with the similarity and implicative relationships of the
variables, with only a number of minor deviations.
Figure 4 illustrates the hierarchical similarity diagram of the “condensed”
variables corresponding to grade 9 students’ responses to the tasks of the
two tests. In line with the general similarity diagram for the whole sample,
two distinct clusters of variables are formed, namely Cluster 1 and 2, which
148 I. Elia and A. Gagatsis

correspond to Test A and B respectively, indicating lack of consistency in

ninth graders’ performance between conversions with a different source rep-
resentation.
Within Cluster 1 two similarity groups are formed. Group 1a involves the
variables of the conversions of a relation represented as a region of points,
xy > 0 (v212a) and of a constant function y = 3/2, represented as a horizon-
tal line parallel to the axis xx0 (v512a). The transformations of these relations
were approached differently from the transformations of the functions y = x−2
(V612a) and y = −x (v412a) and the algebraic relation incorporating a func-
tion y > x (V312a). The transformations of the three latter relations were
tackled similarly due to their common functional character, thus establishing
a group, namely Group 1b. Despite their common properties with the constant
function, the conversion of this relation (v512a) was carried out differently,
indicating students’ difficulty to realize that the graph of a horizontal line
represents a function and handle it as such. This finding deviates from the
second prediction stating that close similarity relationships would be formed
among the variables, which correspond to conversions of functions. Explaining
verbally what the graph y < 0 (v112a) represented was carried out in a dif-
ferent way from the graphs of the other relations of the test probably because
of its distinct perceptual characteristics, which made its interpretation easier.
In Cluster 2 two similarity groups are also identified. Group 2a involves the

Fig. 4. The hierarchical similarity diagram among the responses of students of grade
9 to the conversion tasks of Test A and Test B (C.H.I.C. 3.5)

responses to the conversions of the verbal representation of two algebraic rela-

tions corresponding to regions of points (v112b and v212b). Group 2b, which
presents a stronger similarity than Group 2a, is comprised of the responses
 A comparison between the hierarchical clustering of variables 149

to the conversions of the three functions of the test (v412b, v512b, v612b)
and the algebraic relation which involves a function (v312b: y > x). Thus, the
conceptual components of the algebraic relations, distinguished by whether
they involve a function or not, seem to differentiate students’ processes in the
conversions of a verbal representation (Test B).
Figure 5 illustrates the implicative diagram of the “condensed” variables
corresponding to grade 9 students’ responses to the tasks of the two tests.
The results of the implicative analysis are in line with the similarity relations
explained above. Two separate “chains” of implicative relations among the
variables are formed with respect to the test they refer to, namely Chain A
and Chain B. This suggests that success at the conversions of the algebraic
relations in the two tests depended primarily on their initial representation.
The commonality of their content did not have a role. For instance, students
who succeeded at the conversion of a function with the graphic form as the
source representation did not necessarily succeed at the conversion of the same
function with the verbal form as the initial representation. Students carried
out the conversions by activating compartmentalized processes based on the
source representation of the conversion.
The two implicative chains have a similar structure, which stems from
the conceptual components of the tasks. In particular, the conversions of the
functions y = x − 2 (v612a or v612b) and y = −x (v412a or v412b) were the
most difficult tasks, and success at them implied success at the conversions
of almost all the other algebraic relations in each test. The conversions of
the constant function (v512a or v512b) and the algebraic relation involving a
function (v312a or v312b) were less complex. Students exhibited the greatest
facility at the conversion tasks of the algebraic relations standing for regions
of points (y < 0 or xy > 0). Figure 6 illustrates the hierarchical similarity
diagram of the variables corresponding to grade 11 students’ responses to the
tasks of the two tests. The structure of the similarity relations in this diagram
is analogous to the structure in the diagram concerning grade 9 students, as
two similarity clusters are established with respect to the source represen-
tation of the conversions. A main difference though is that the first cluster
(Cluster 1), which refers to the conversions of graphic representations, involves
one similarity group (Group 1a), in which the variable of the fifth task (v512a)
is linked to the variables of the third (v312a), fourth (v412a) and sixth (v612a)
task. Thus, students carried out the conversion of the constant function us-
ing similar processes with the ones when performing the conversions of the
other algebraic relations involving functions. This similarity is weaker though
than the similarities among the conversions of the other relations (3, 4 and 6).
Eleventh graders’ increased consistency in comparison with the ninth graders’
consistency indicates the older students’ realization that the graph of the re-
lation y = 3/2 represents a function despite its dissimilar perceptual form
relatively to the graphs of the other relations involving functions. However,
students performed the conversions of the other relations, i.e. y < 0 (v112a
150 I. Elia and A. Gagatsis

Fig. 5. The implicative diagram among the responses of students of grade 9 to the
conversion tasks of Test A and Test B (C.H.I.C. 3.5)

or v112b) and xy > 0 (v212a or v212b), starting from a verbal or a graphic

form in a different way, probably because they did not represent functions.

Fig. 6. The hierarchical similarity diagram among the responses of students of grade
11 to the conversion tasks of Test A and Test B (C.H.I.C. 3.5)
 A comparison between the hierarchical clustering of variables 151

Figure 7 shows the implicative diagram of the variables corresponding to

grade 11 students’ responses to the tasks of the two tests. This implicative di-
agram is congruent with the implicative diagram concerning grade 9 students.
Specifically, a basic commonality is the establishment of two distinct chains
of implicative relations, among the variables with respect to the type of con-
version or the test they refer to, namely Chain A and Chain B. Furthermore,
like in the ninth graders’ similarity diagram, success at the conversions of the
algebraic relations involving functions implied success at the conversions of
the relations corresponding to regions of points in each test.

Fig. 7. The implicative diagram among the responses of students of grade 11 to the
conversion tasks of Test A and Test B (C.H.I.C. 3.5)

However, in this diagram, the structure of each implicative chain has a

more linear form relatively to the diagram of ninth graders, illustrating more
explicitly the hierarchical ordering of the conversion tasks with respect to the
type of the algebraic relation involved. In particular, the most difficult tasks
which implied success to all the other tasks in each test involved the conver-
sion of the function y = x − 2 (v612a or v612b) (see Appendix). Less complex
152 I. Elia and A. Gagatsis

conversion tasks in the two tests involved the function y = −x (v412a or

v412b) and students who tackled them succeeded at the easier tasks incorpo-
rating the algebraic relation y > x (v312a or v312b). Consecutively, students
who carried out the latter tasks were successful at the conversion tasks of
the constant function y = 3/2 (v512a or v512b) in both tests. In Chain B
(Test B), success at the conversion task of the verbal representation of y > x
implied directly success also at the conversions of another verbally given in-
equality, that is, xy > 0 (v212b), which in turn entailed success at the easiest
conversion task of the test, involving the inequality y < 0 (v112b). In Chain
A (Test A), carrying out the conversion task of the graphic representation of
y = 3/2 (v512a) implied success at the conversions of the relation xy > 0
(v212a) corresponding to a region of points, which in turn entailed success-
ful performance in the conversions of the graphic form of the relation y < 0
(v112a) representing a region of points as well.

6 Discussion

This study investigated students’ abilities in the conversions of functions and

other algebraic relations among representations, as well as, their interrela-
tions. The data, which were collected using two tests involving conversion
tasks of the same mathematical content with a different source representa-
tion, were analyzed from different perspectives using three distinct statistical
methods, each of which is based on a different rationale. CFA is used to test
statistically whether a hypothesized connection pattern between the observed
variables and their underlying factors exists. The a priori hypothesis in this
study stemmed from knowledge of past empirical work and theory which sug-
gested that students encounter difficulties in transferring information gained
in one representational context to another related to the concept of func-
tions [1, 3, 9, 10]. The hierarchical clustering of variables aims at bringing to
light the consistency among students’ responses to the various tasks in a hi-
erarchical manner. The implicative method gives information about whether
success at one task implies success at another task and about the relative dif-
ficulty of the tasks based on students’ performance (Table 2, item 1). A major
concern of this study is to compare in detail the findings of these statistical
analyses so as to gain insight about the advantages of each method, as well
as, about whether their outcomes on the same sample data are congruent and
can complement each other.
Tables 2 and 3 summarize and compare the key outcomes of the three
statistical methods on the data of the study that concur or have a com-
plementary role amongst them. Whereas the implicative technique and the
hierarchical clustering of variables incorporated only observed measurements,
CFA allowed the development and validation of a model that involves not only
observed but also latent (unobserved) variables, which can not be observed
or measured directly (Table 2, item 2). These constructs, which lied behind
 A comparison between the hierarchical clustering of variables 153

the corresponding observed measures, were the ability to perform conversions

of algebraic relations from the graphic representation to symbolic and to ver-
bal form, and the ability to carry out conversions of algebraic relations from
the verbal representation to symbolic and to graphic form. The fact that two
factors are needed to account for the effects of the two types of the initial
represantation exemined here, i.e. graphic form and verbal form provided a
strong case for the role of the source represantation in the conversions of alge-
braic relations ammong different representations. Another abstract construct
of a higher-order level was assumed to underlie these abilities, indicating that
despite the discrepancy in students’ performance among the different types of
conversion, both abilities are still basic components of a common construct,
i.e. general ability in transferring an algebraic relation from one representation
to others.
The difference in the strength of the relations of the two first-order factors
to the second-order factor in the model revealed that the graphic and the
verbal representations have a different and an almost autonomous function in
the conversions of algebraic relations. The outcomes of the other two methods
of analysis were the ones that revealed in a more explicit way students’ com-
partmentalized responses to the conversion tasks with respect to the source
representation (Table 2, item 3). The separate grouping of the responses to
conversions having as the initial representation the verbal form or the graphic
form in the similarity diagrams showed students’ inconsistency when dealing
with conversions of different initial representations. The implicative diagrams
included distinct chains of variables with respect to the initial representation of
the conversion indicating that success in one type of conversion of an algebraic
relation did not necessarily imply success in another mode of conversion of the
same relation. Weak connections or lack of implications among conversions of
the same mathematical content with a different starting representation are the
main features of the phenomenon of compartmentalization and indicate that
students did not construct the whole meaning of the concept of function and
did not grasp the whole range of its applications [10]. As Even supports [16],
the ability to identify and represent the same concept in different represen-
tations, and flexibility in moving from one representation to another allow
students to see rich relationships, and develop a deep understanding of the
concept.
In comparing the three methods of analysis, the CFA validated the same
grouping of tasks as the implicative method and the hierarchical classification
method, since the measurement indicators of each factor in the model formed
a separate group in the other two analyses. Nevertheless, the implicative and
the hierarchical classification methods provided further insight and a more
analytic view about the construction and hierarchical structure of these groups
and the implicative relations among students’ responses to the tasks (Table 4).
For example, the hierarchical clustering of variables revealed some discrep-
ancies in the ways students tackled particular tasks of the two tests, that were
not evident in the CFA model (Table 4, item 1). Whereas the conversions of
154 I. Elia and A. Gagatsis

CFA Hierarchical cluster- Implicative method

ing
1. Factorial structure of 1. Hierarchical classifi- 1. Implicative rela-
students’ abilities in the cation and consistency tions between students’
conversion of algebraic of students’ responses to responses to the conver-
relations from one repre- the conversions sions, relative difficulty
sentation to another of the tasks
2. Development of a 2. Similarity groupings 2. Implications among
model involving two among observed mea- observed variables
latent (unobserved) surements standing for standing for students’
factors for the effects students’ responses to responses to the con-
of two types of initial the conversions of al- versions of algebraic
representation and a gebraic relations having relations having the
second-order factor the verbal or the graphic verbal or the graphic
standing for the gen- mode as the source rep- mode as the source
eral ability to convert resentation representation
functions from one
representation to others
3. Difference in the 3. Compartmental- 3. Two implicative
strength of the relations ization in students’ chains involving stu-
of the two first-order fac- responses to the con- dents’ responses to the
tors to the second-order versions with respect to conversions of a verbal
factor: the graphic and their source representa- representation and the
the verbal form of the tion conversions of a graphic
same content operate one respectively, lack
rather autonomously in of implications between
the conversions the variables of the two
chains
4. Lower factor loadings 4. Separate grouping of 4. The conversions in-
of the conversions of in- the variables of conver- volving functions were
equalities relatively to sions of inequalities and more complex than the
the conversions of func- the variables of conver- tasks involving regions of
tions: considerable au- sions of functions: incon- points. Success at carry-
tonomy among the con- sistency between them, ing out a conversion of
version processes of func- significant role of the a function implied suc-
tions and the conversions conceptual properties of cess at a conversion of
of algebraic relations not the algebraic relations on the same type involving
representing functions students’ consistency an inequality
Table 2. The congruent and complementary outcomes of the CFA, the hierarchical
clustering of variables and the implicative method on the data of the study (Part 1)
 A comparison between the hierarchical clustering of variables 155

5. Lower factor load- 5. Relatively weak sim- 5. The conversion of the

ing of the conversion of ilarity of students’ con- constant function was
the graphic form of a versions of the graphic the easiest one among
constant function rela- form of the constant the conversion tasks of
tively to the correspond- function and of the other the other functions.
ing conversions of the functions: distinct ways
other functions of approaching the con-
version of a graphic rep-
resentation of this kind
of function
6. The factorial structure 6. The abilities to 6. Two distinct chains
remains invariant across carry out conversions with respect to the type
grades. of graphic or verbal of the conversions in
representations of alge- both grades, success at
braic relations remain a conversion of a graph
compartmentalized in does not imply success at
the two grades a conversion of a verbal
representation in neither
grade.
Table 3. The congruent and complementary outcomes of the CFA, the hierarchical
clustering of variables and the implicative method on the data of the study (Part 2)

the verbal representations of the algebraic relations y < 0 and xy > 0 were
approached similarly to each other, this was not the case in the conversions
of the graphic representations of the corresponding relations. This outcome
provides evidence to the students’ disconnected and distinct ways of using the
verbal form and the graphic form as source representations in conversions,
giving further support to the phenomenon of compartmentalization. The im-
plicative diagram indicated that the conversion of the graphic representation
of the function y = x − 2 (v612a) was the most complex task of both tests
and students who provided a correct solution at it, succeeded at all of the
other conversion tasks of both tests. Moreover, the implicative relations link-
ing responses to tasks of the two tests showed that success at some tasks of
Test A entailed success at the corresponding tasks of Test B. The above out-
comes of the implicative method provide evidence to the distinct and more
difficult character of the conversions starting with a graph relatively to con-
versions starting with a verbal description. This differentiation and increased
difficulty may be due to the fact that the perceptual analysis and synthesis
of mathematical information presented implicitly in a diagram often make
greater demands on a student that any other aspect of a problem [22]. The
graphic register functions effectively only under the conventions of a different
mathematical culture. Due to students’ poor knowledge of this new culture,
graphic modes of representation are difficult to decode and work out. It can
be asserted that the support offered by mathematical meta-language is more
156 I. Elia and A. Gagatsis

fundamental than the aid given by the graphic register for carrying out a
translation from one mode of representation to another [10].
Furthermore, the outcomes of the three statistical processes uncovered how
students dealt with different types of algebraic relations in conversions of the
same source representation (Tables 2 and 3, items 4 and 5). The conversions
of the algebraic relations that corresponded to inequalities and thus regions
of points were found to have considerable autonomy from the conversions of
relations involving functions. The separate grouping of the former variables
from the latter ones in the similarity diagrams revealed that students tackled
the conversions of inequalities differently from the conversions of functions.
The lower factor loadings of the variables referring to the conversions of in-
equalities relatively to the corresponding loadings of the variables standing for
the conversions of functions in the CFA model were in line with this outcome.
The implicative diagrams revealed additional information to the above find-
ings, suggesting that the tasks involving functions were more complex than
the tasks involving regions of points in any type of conversion. Moreover, the
implications among the variables showed that students who carried out con-
versions of function from one representation to another were able to succeed
at conversions of the same type involving inequalities. Considering the rela-
tions within the responses to the function tasks, a relatively weak similarity
was observed between students’ response to the conversion of the constant
function and their responses to the other functions, revealing their distinct
ways of approaching the conversion of a graphic representation of this kind
of function. Given that this distinction did not apply in the conversions of
the verbal representation (Test B), it is indicated that the interpretation of
the graphic form of the particular type of function was the main factor dif-
ferentiating students’ performance. The relatively lower factor loading of the
variable referring to the conversion of a constant function from a graphic rep-
resentation to another representation in comparison to the factor loadings of
the corresponding variables of the other functions gave further support to
students’ different conversion processes when dealing with this special kind of
function. The implicative relations revealed that the task involving the par-
ticular function was the easiest one among the conversion tasks of the other
functions. The above findings highlight the effect of the inherent mathemat-
ical foundation of the algebraic relations, i.e. whether they are functions or
not and what kind of functions they are, on students’ processes, consistency
and success in conversions of the same type.
Addressing the effect of age on the abilities to transfer algebraic rela-
tions from one representation to another, the outcomes of all of the analyses
(Table 3, item 6) indicated that students’ compartmentalized ways of using
representations and thinking of algebraic relations in general and functions in
particular occurs in both grades. Specifically, the factorial structure described
above remains invariant, while the abilities to carry out conversions of graphic
or verbal representations of algebraic relations remain compartmentalized in
the similarity and the implicative diagrams for the two grades involved in the
 A comparison between the hierarchical clustering of variables 157

study. These findings, which provide support to the results of the studies by
Elia et al [10] and Gagatsis and Christou [11] indicate that relations among
translations from one mode of representation to another do not vary as a func-
tion of grade levels. Thus the relative inherent nature of difficulties of each
type of conversion has age endurance, suggesting that development or regular
instruction does not change students’ processes while dealing with conversions
of functions from one representation to another.
Despite this invariance, some discrepancies in the performance of students
of the two grades occurred as regards the inherent mathematical properties
of the algebraic relations (Table 4, item 2). Based on the similarity diagrams,
eleventh graders were found to respond to the conversion of the graph of the
constant function more coherently with the conversions of the other functions
relatively to the ninth graders. This suggests that the older students were
more competent in recognizing the conceptual components of some algebraic
relations, i.e. whether they represented a function or not, and dealt with
their conversions more consistently. This competence was also indicated by
the outcomes of the implicative analysis on the data of the two age groups
respectively. The hierarchical ordering of the tasks with respect to their level of
difficulty was clearer in the implicative diagram of eleventh graders compared
to the diagram of the ninth graders. Given that the different levels of difficulty
of the tasks stemmed from the type of the algebraic relations involved, it
can be asserted that the older students identified more efficiently the distinct
conceptual properties of each algebraic relation. Thus, the conceptual features
of the algebraic relations seemed to have a stronger impact on their processes
and success levels relatively to the younger students’ responses.

In general, the application of all of the analyses yielded congruent results.

However, at the same time given that these statistical processes approached
the data from different perspectives, they emphasized different aspects of stu-
dents’ outcomes. This differentiation allowed for the accumulation of a number
of new distinctive elements by each analysis that contributed to the unravel-
ling and making sense of students’ performance, structure of abilities, difficul-
ties and inconsistencies on the particular subject. The findings of the study
suggest that the three statistical methods are open to complementary use
and each one does not operate at the expense of the other. CFA provided a
means for making sense of the structure of students’ abilities in the conver-
sion of functions among different representations. The hierarchical clustering
of variables provided a means for classifying students’ responses, for identi-
fying students’ consistencies and inconsistencies among different conversions
and for investigating the factors influencing this behaviour. The implicative
method provided a means for examining the implicative relations among the
responses to the tasks and the relative difficulty of the different conversions on
the basis of students’ performance. Provided that applying these methods of
analysis is consistent with the objectives of a study, their combination on the
same sample data could contribute to overcome some significant limitations of
158 I. Elia and A. Gagatsis

Hierarchical clustering Implicative method

1. The conversions of the verbal rep- 1a) The conversion of the graphic rep-
resentations of the algebraic relations resentation of the function y = x − 2
y < 0 and xy > 0 were approached simi- was the most complex task of both tests;
larly to each other unlike the conversions b) Success at the tasks of Test A en-
of the graphic representations of the cor- tailed success at the corresponding tasks
responding relations: disconnected ways of Test B: greater difficulty of the con-
of using the verbal form and the graphic version starting with a graph relatively
form as source representations to a conversion starting with a verbal de-
scription
2. The conversions of the constant func- 2. The hierarchical ordering of the tasks
tion and of the other functions were tack- with respect to their level of difficulty
led more coherently by the 11th graders was clearer in the 11th graders’ diagram
than by the 9th graders: the older stu- compared to the 9th graders’ diagram:
dents were more competent in recogniz- the older students were more able to
ing the conceptual components of some identify the distinct conceptual proper-
algebraic relations ties of each algebraic relation

Table 4. The new additional outcomes of the hierarchical clustering of variables

and the implicative method to the outcomes of the CFA

each analysis employed separately, and consequently could enrich and deepen
the outcomes of the investigation.

1
Glossary of initials used in the text:
CFA: Confirmatory Factor Analysis
EQS: Bentler’s Structural Equation Modeling program
CFI: Comparative Fit Index
RMSEA: Root Mean Square Error of Approximation
SEM: Structural Equation Modeling
EFA: Exploratory Factor Analysis
CHIC: Classification Hierarchique, Implicative et Cohesitive

References
1. R. Duval. The cognitive analysis of problems of comprehension in the learning
of mathematics. Mediterranean Journal for Research in Mathematics Education,
1(2):1–16, 2002.
2. J. Kaput. Representation systems and mathematics. In C. Janvier (ed.): Prob-
lems of representation in the teaching and learning of Mathematics, pages 19–26,
Lawrence Erlbaum Associates Publishers, Hillsdale NJ, 1987.
3. A. Gagatsis, M. Shiakalli. Ability to translate from one representation of the
concept of function to another and mathematical problem solving. Educational
Psychology, 24(5):645–657, 2004.
 A comparison between the hierarchical clustering of variables 159

4. A. Sierpinska. On understanding the notion of function. In E. Dubinsky, G. Harel

(eds.): The concept of function: Aspects of epistemology and pedagogy, pages 25–28,
The Mathematical Association of America, United States, 1992.
5. R. Lesh, T. Post, M. Behr. Representations and translations among represen-
tations in mathematics learning and problem solving. In C. Janvier (Ed.): Prob-
lems of representation in the teaching and learning of Mathematics, pages 33–40,
Lawrence Erlbaum Associates Publishers, Hillsdale NJ, 1987.
6. A. Gagatsis, I. Elia, A. Mougi. The nature of multiple representations in devel-
oping mathematical relations. Scientia Paedagogica Experimentalis, 39(1):9–24,
2002.
7. A. Evangelidou, P. Spyrou, I. Elia, A. Gagatsis. University students’ conceptions
of function. In M.Jonsen Hoines, A. Berit Fuglestad (eds.): Proceedings of the
28th Conference of the International Group for the Psychology of Mathematics
Education, pages 351–358, Bergen University College, Bergen, Norway, 2004.
8. I. Elia, A. Panaoura, A. Eracleous, A. Gagatsis. Relations between secondary
pupils’ conceptions about functions and problem solving in different representa-
tions. International Journal of Science and Mathematics Education, 5:533–556,
2007.
9. I. Elia, P. Spyrou. How students conceive function: A triarchic conceptual-
semiotic model of the understanding of a complex construct. The Montana Math-
ematics Enthousiast 3(2):256–272, 2006.
10. I. Elia, A. Gagatsis, R. Gras. Can we “trace” the phenomenon of compart-
mentalization by using the I.S.A.? An application for the concept of function. In
R. Gras, F. Spagnolo, J. David (eds.): Proceedings of the Third International Con-
ference I.S.A. Implicative Statistic Analysis, pages 175–185, Universita degli Studi
di Palermo, Palermo, Italy, 2005.
11. A. Gagatsis, C. Christou. The structure of translations among representations
in functions. Scientia Paedagogica Experimentalis, 39(1):39–58, 2002.
12. B. Dufour-Janvier, N. Bednarz, M. Belanger. Pedagogical considerations con-
cerning the problem of representation. In C. Janvier (ed.): Problems of repre-
sentation in the teaching and learning of mathematics, pages 109–122, Lawrence
Erlbaum Associates Publishers, Hillsdale NJ, 1987.
13. J. G. Greeno, R.P. Hall. Practicing representation: Learning with and about
representational forms. Phi Delta Kappan, 78:361–367, 1997.
14. S. Ainsworth, P. Bibby, D. Wood. Evaluating principles for multi-
representational learning environments. Paper presented at the 7th European Con-
ference for Research on Learning and Instruction, Athens, Greece, 1997.
15. J. Kaput. Technology and mathematics education. In D. A. Grouws (ed.): Hand-
book of research on mathematics teaching and learning, pages 515–556, Macmillan,
New York, 1992.
16. R. Even. Factors involved in linking representations of functions. The Journal
of Mathematical Behavior, 17(1): 105–121, 1998.
17. R. Duval. A cognitive analysis of problems of comprehension in a learning of
mathematics. Educational Studies in Mathematics, 61(1–2):103–131, 2006.
18. J.P. Smith, A.A. diSessa, J. Roschelle. Misconceptions reconceived: A con-
structivist analysis of knowledge in transition. Journal of the Learning Sciences,
3:115–163, 1993.
19. T. Eisenberg, T. Dreyfus. On the reluctance to visualize in mathematics. In
W. Zimmermann, S. Cunningham (eds.): Visualization in Teaching and Learning
160 I. Elia and A. Gagatsis

Mathematics, pages 9–24, Mathematical Association of America, United States,

1991.
20. A. Sfard. Operational origins of mathematical objects and the quandary of reifi-
cation — The case of function. In E. Dubinsky, G. Harel, (eds.): The concept of
function: Aspects of epistemology and pedagogy, pages 59–84, The Mathematical
Association of America, United States, 1992.
21. Z. Markovits, B. Eylon, M. Bruckheimer. Functions today and yesterday. For
the Learning of Mathematics, 6(2):18–28, 1986.
22. L. Aspinwall, K. L. Shaw, N. C. Presmeg. Uncontrollable mental imagery:
Graphical connections between a function and its derivative. Educational Stud-
ies in Mathematics, 33:301–317, 1997.
23. R. Duval. Registres de Représentation Sémiotique et Fonctionnement Cognitif
de la Pensée. Annales de Didactique et de Sciences Cognitives, 5:37–65, 1993.
24. B. M. Byrne. Structural Equation Modeling with EQS and EQS/Windows: Basic
concepts, applications and programming, SAGE Publications Inc., Thousand Oaks
CA, 1994.
25. P. M. Bentler. Causal modeling via structural equation systems. In
J. R Nesselroade, R. B. Cattell (eds.): Handbook of multivatriate experimental
psychology, (2nd ed.), pages 317–335, Plenum, New York, 1988.
26. R. B. Kline. Principles and practice of structural equation modeling, Guilford
Press, New York, 1998.
27. P. M. Bentler. EQS structural equations program manual, Multivariate Software
Inc., Encino CA, 1995.
28. P. M. Bentler. Comparative fit indexes in structural models. Psychological Bul-
letin, 107:301–345, 1990.
29. A. Bodin, R. Coutourier, R. Gras. CHIC: Classification Hiérarchique Implicative
et Cohésitive-Version sous Windows, CHIC 1.2. Association pour la Recherche en
Didactique des Mathématiques, Rennes, 2000.
30. R. Gras. Data analysis: a method for the processing of didactic questions. Se-
lected papers for ICME 7, La Pensée Sauvage, Grenoble, 1992.
31. R. Gras, et al. L’implication statistique, Collection associée à Recherches en
Didactique des Mathématiques, La Pensée Sauvage, Grenoble, 1996.
32. I.C. Lerman. Classification et analyse ordinale des données, Dunod, Paris, 1981.
33. R. Gras, P. Peter, H. Briand, J. Philippe. Implicative Statistical Analysis.
In C. Hayashi, N. Ohsumi, N. Yajima, Y. Tanaka, H. Bock, Y. Baba (eds.):
Proceedings of the 5th Conference of the International Federation of Classifica-
tion Societies, pages 412–419, Springer-Verlag, Tokyo, Berlin, Heidelberg, New
York, 1997.
 A comparison between the hierarchical clustering of variables 161

Appendix

Test A
N=183 Graphic→Verbal Graphic→Symbolic
Occurrence Mean Occurrence Mean
V1: y <0 139 .76 102 .56
V2: xy > 0 93 .51 72 .39
V3: y > x 67 .37 46 .25
V4: y = −x 75 .41 36 .20
V5: y = 3/2 65 .36 70 .38
V6: y = x − 2 28 .15 13 .07
Test B
N=183 Verbal→Graphic Verbal→Symbolic
Occurrence Mean Occurrence Mean
V1: y <0 132 .72 109 .60
V2: xy > 0 116 .63 72 .39
V3: y > x 83 .45 91 .50
V4: y = −x 58 .32 47 .26
V5: y = 3/2 68 .37 80 .44
V6: y = x − 2 53 .29 44 .24

Table 5. Frequencies of occurrence and means of Grade 9 students to the tasks of

Test A and Test B
162 I. Elia and A. Gagatsis

Test A
N=404 Graphic→Verbal Graphic→Symbolic
Occurrence Mean Occurrence Mean
V1: y < 0 322 .80 288 .71
V2: xy > 0 244 .60 193 .48
V3: y > x 171 .42 145 .36
V4: y = −x 166 .41 137 .34
V5: y = 3/2 205 .51 177 .44
V6: y = x − 2 156 .39 107 .26
Test B
N=404 Verbal→Graphic Verbal→Symbolic
Occurrence Mean Occurrence Mean
V1: y < 0 329 .81 300 .74
V2: xy > 0 339 .84 218 .54
V3: y > x 250 .62 269 .67
V4: y = −x 195 .48 224 .55
V5: y = 3/2 257 .64 286 .71
V6: y = x − 2 204 .50 190 .47

Table 6. Frequencies of occurrence and means of Grade 11 students to the tasks of

Test A and Test B
Implications between learning outcomes
in elementary bayesian inference

Carmen Díaz1 , Inmaculada de la Fuente2 and Carmen Batanero3

1
Facultad de Psicología, Campus El Carmen, Universidad de Huelva
21071 Huelva, Spain
[email protected]
2
Facultad de Psicología, Campus de Cartuja, Universidad de Granada
18071 Granada, Spain
[email protected]
3
Facultad de Educación, Campus de Cartuja, Universidad de Granada
18071 Granada, Spain
[email protected]

Summary. In this research implicative analysis served to study some previous hy-
potheses about the interrelationships in students’ understanding of different con-
cepts and procedures after 12 hours of teaching elementary Bayesian inference. A
questionnaire made up of 20 multiple choice items was used to assess learning of
78 psychology students. Results suggest four groups of interrelated concepts: con-
ditional probability, logic of statistical inference, probability models and random
variables.

Key words: Bayesian inference, teaching, conditional probability, undergraduate

students, assessment

1 Introduction

There is a tendency nowadays to recommend that teaching of Bayesian infer-

ence should be included in undergraduate statistics courses as an adequate
and desirable complement to classical inference [22, 25, 26]. Situations where
prior information can help to make an accurate decision and software that fa-
cilitates the application of these methods are becoming increasingly available.
Moreover, top core statistical journals now include an important proportion
of Bayesian papers but this does not yet translate into comparable changes in
the teaching of statistical inference to undergraduates [6].
Some excellent textbooks whose understanding does not involve advanced
mathematical knowledge and where basic elements of Bayesian inference are
contextualized in interesting examples (e.g. [7] or [2]) can help follow these
recommendations. There are also a wide number of Internet didactic resources
C. Díaz et al.: Implications between learning outcomes in elementary bayesian inference,
Studies in Computational Intelligence (SCI) 127, 163–184 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
164 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

that might facilitate the teaching of these concepts (e.g. those available from
Jim Albert’s web page at http://bayes.bgsu.edu/). These and other authors
(e.g. [8]) have incorporated Bayesian methods to their teaching and are sug-
gesting that Bayesian inference is easier to understand than classical inference.
This is however a controversial point. On one hand, it is argued [29] that
Bayesian inference relies too strongly on conditional probability, a topic hard
for undergraduate students in non-mathematical majors to learn. On the other
hand, in the past 50 years errors and difficulties in understanding and apply-
ing frequentist inference have widely been described (e.g. in [3, 21]). These
criticisms suggest researchers do not fully understand the logic of frequentist
inference and give a (incorrect) Bayesian interpretation to p-values, statis-
tical significance and confidence intervals. It is then possible that learning
Bayesian inference is not as intuitive as assumed or at least that not all the
concepts involved are equally easy for students. Moreover, empirical research
that analyze the learning of students in natural teaching contexts is almost
non-existent.
Consequently, the first aim of this research was to explore the extent to
which different concepts involved in basic Bayesian inference are accessible
to undergraduate psychology students. A second goal was to compare learn-
ing outcomes with our previous hypotheses that there are different groups of
related concepts (and not just conditional probability) that are potentially
difficult for these students. We finally wanted to explore the implications be-
tween concepts included in each of these groups with the aim of providing
some recommendations about how to best organize the teaching of the topics.
In this sense, implicative analysis was an essential tool. As suggested
by [17], researchers in human sciences are interested in discovering inductive
nonsymmetrical rules of the type “if a, then almost surely b”. The method
provides an implication index for different types of variables; and moreover
serves to represent these implications in a graph or an implicative hierarchy
as a complex non linear system. This specially suits our theoretical frame-
work [16], where knowledge is seen as a complex system, more than as a
linear object, and for this reason, in our research we were interested in finding
the implications between understanding the different mathematical objects
involved in basic Bayesian inference, that is, what the knowledge contents A
that facilitate learning of other different contents B are.

2 Teaching Experiment
The sample taking part in this research included 78 students (18–20 years
old) in the first year of the Psychology Major at the University of Granada,
Spain. These students were taking part in the introductory statistics course
and volunteered to take part in the experiment. The sample was composed of
17.9% boys and 82.1% girls, which is the normal proportion of boys and girls
 Implications between learning outcomes in elementary bayesian inference 165

in the Faculty. These students scored an average of 4.83 (in a scale 0–10) in
the statistics course final examination with standard deviation of 2.07.
The students were organized into four groups of about 15–20 students each
and attended a short 12 hour long course given by the same lecturer with the
same material. The 12 hours were organized into 4 days. Each day there were
two teaching sessions with a half hour break in between. The first session (2
hours) was dedicated to presenting the materials and examples, followed by
a short series of multiple choice items that each student should complete, in
order to reinforce their understanding of the theoretical content of the lesson.
In the second session (one hour), students in pairs worked in the computer
lab with the following Excel programs that were provided by the lecturer to
solve a set of inference problems:
1. Program Bayes: This program computes posterior probabilities from prior
probabilities and likelihood (that should be identified by the students from
the problem statement).
2. The program Prodist transforms a prior distribution P (p = p0 ) for a
population proportion p in the posterior distribution P (p = p0 | data),
once the number of successes and failures in the sample are given. Prior
and posterior distributions are drawn in a graph.
3. The program Beta computes probabilities and critical values for the Beta
distribution B(s, f ), where s and f are the numbers of successes and
failures in the sample.
4. The program Mean computes the mean and standard deviation in the
posterior distribution for the mean of a normal population, when the mean
and standard deviation in the sample and prior population are known.
In Table 1 we present a summary of the teaching content. Students were
given a printed version of the didactic material that covered this content. Each
lesson was organized in the following sections: a) Introduction, describing the
lesson goals and introducing a real life situation; b) Progressive development
of the theoretical content, in a constructive way and using the situation pre-
viously presented; c) Additional examples of other applications of the same
procedures and concepts in other real situations, d) Some solved exercises,
with description of main steps in the solving procedure; e) New problems that
students should solve in the computer lab; and f) Self assessment items. All
this material together with the Excel programs described above was also made
available to the students on the Internet (http://www.ugr.es/ mcdiaz/bayes).
We added a forum, so that students could consult the teacher or discuss them-
selves their difficulties, when needed.

3 A-Priori Analysis of the Questionnaire

Two weeks after the end of the teaching, the students were given a question-
naire to assess their understanding of the topic. Students prepared in advance
166 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

Lesson Content Session 1: classroom Session 2: computer lab.

1 Bayes theorem in the Prior and posterior Solving Bayes problems
context of clinical di- probabilities; likelihood; (Program Bayes)
agnosis Bayes theorem;
comparing subjective
and frequentist
probability; revision of
beliefs; sequential
application of Bayes
theorem
2 Inference for propor- Parameters as random Computing credible
tion. Discrete case in variables; prior and pos- intervals for proportion;
the context of voting terior distribution; in- assigning non informa-
formative and non in- tive and informative
formative prior distrib- prior distributions
ution; credible intervals; (Program Prodist)
comparing Bayesian and
frequentists approaches
to inference
3 Inference for propor- Generalizing to continu- Assigning non informa-
tion. Continuous case ous case; Beta distribu- tive and informative
in the context of pro- tion; its parameters and prior distributions; com-
duction shape; credible intervals; puting credible intervals
Bayesian tests for proportion; test-
ing simple hypotheses
(Program Beta)
4 Inference for the Normal distribution and Assigning non informa-
mean of a normal its parameters; credible tive and informative
population in the intervals and tests for prior distributions; com-
context of psycholog- the mean of a nor- puting credible intervals
ical assessment mal distribution with for means; testing simple
known variance; non in- hypotheses (Program
formative and informa- Mean)
tive prior distributions
Table 1. Teaching content and its organization

for the assessment that was part of the analysis course they were following.
The BIL (Bayesian Inference Learning) questionnaire (which is included in
Appendix) was prepared for this research and is composed of both multiple
choice and some open ended items that were developed by the authors with
the specific aim to cover the most important contents in the teaching. The aim
was to assess learning in the following groups of concepts, which in our a-priori
analysis were assumed to be the core content of basic Bayesian inference and
might cause different types of difficulties to students. These concepts, as well
as the philosophical principles of Bayesian inference had been introduced in
the teaching at an elementary level, adequate to the type of students. We also
assumed learning of one of these groups of concepts would not automatically
 Implications between learning outcomes in elementary bayesian inference 167

assure the learning of the other groups, so in the implicative analysis the three
groups could be unrelated.

Conditional probability and the Bayes’ theorem

As was argued before, different authors pointed to students difficulties in

understanding conditional probability. For example, the students’ confusion
between the two probabilities P (A | B) and P (B | A) was termed the fallacy of
the transposed conditional [13]. [20, 34] described the identification of causal-
ity and conditioning (causal conception of conditional probability) and the
belief that an event could not condition another event that occurs before it
(chronological conception); confusion between simple, joint and conditional
probability was described by [31]. All these errors might cause difficulties in
computing different types of probabilities (item 2), understanding of the dif-
ferences between prior and posterior probability and likelihood (items 1 and
18), and using the Bayes’ theorem as a tool to transform prior into posterior
probabilities (item 7 and 18). In addition, students’ difficulties with the Bayes’
theorem were also described by the afore mentioned and other authors (see [5]
for a survey).

Parameters as random variables, their distribution, distinction

between prior and posterior distribution

In Bayesian inference, parameters are considered to be random variables with

a prior distribution, while in frequentist inference they are assumed to be
unknown constants (items 3, 5), a distinction which is not too clear for some
students [23]. Moreover, the aim of Bayesian inference is to transform the
prior into a posterior distribution via the Bayes’ theorem (item 18). A prior
distribution provides all the information for the parameter before collecting
the data (item 4), non informative priors are given by uniform distributions
and are used when no previous information is available for the parameter
(item 6).
There are different models to represent prior distributions. The Beta distri-
bution was introduced in the teaching, and students had to learn the meaning
of its parameters (item 8, 20) and how to select a specific Beta distribution
in a particular inference problem (item 9). Students knew the normal distrib-
ution from previous lessons. However, they had to learn the rule to compute
the posterior distribution for a mean when the prior distribution is normal
(item 13, 14, 15, 16). In managing all these distributions, Bayesian statistics
uses the rules of probability to make inferences, and that requires dealing
with formulae, but actual calculus used is minimal as students only have to
understand that probability is given by different types of areas under a density
function [8]. However, the extent to which all of this is grasped by psychology
students has still to be assessed.
168 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

Logic of Bayesian inference

The aim of Bayesian inference is updating the prior distribution via the like-
lihood to get the posterior distribution, which provides all the information
for the parameter, once the data have been collected [9]. However, it is also
possible to carry out procedures similar to those used in frequentist statistics,
although the interpretation and logic is a little different [7, 16].
Credible intervals provide the epistemic probability that the parameter
is included in a specific interval of values, for the particular sample, while
confidence intervals provide the frequentist probability that in a percentage of
samples from the same population the parameter will be included in intervals
of values computed in those samples. Credible intervals are computed from
the posterior distribution (item 17) and students should be able to compute
them by using the tables of different distributions (items 10, 16); they should
understand that the interval width increases with the credibility coefficient
and decreases with the sample size (item 12).
In Bayesian inference we can compare at the same time different hypothe-
ses; in this case we compute the probabilities for those hypotheses given the
data by using the posterior distribution and select the hypothesis with higher
probability (item 11). In testing only one hypothesis we either compute the
probability for the hypothesis or for the contrary event (item 14); acceptance
or rejection will depend on the value of that probability.
So, there are some conceptual and interpretative differences between clas-
sical and frequentist approaches, but, since both approaches often lead to ap-
proximately the same numerical results, students might not understand these
differences and confuse both approaches [23].

4 Implications in Learning Outcomes

In order to assess understanding of the three groups of concepts above and see
if our previous hypothesis that learning of the three groups of concepts might
be unrelated, we gave the BIL questionnaire to the students who participated
in the teaching experiment and analyzed their responses4 . We also were in-
terested in the extent to which each task was easy for the students. In table 2
we present the number and percentage of correct responses to items and sub
items. In item 18 we considered three different scores: correct identification
of prior probabilities (18.1), correct identification of likelihood (18.2) and cor-
rect computation of posterior probabilities (18.3). In item 2 each response was
scored independently.
The average number of correct responses per student was x = 16.6. Given
that the maximum possible score was 26; these results show a reasonable result
4
This assessment was complemented with analysis of self- assessment tests and
open tasks solved by the students along the teaching sessions. We are not including
here the analysis of this complementary data, due to space limitation.
 Implications between learning outcomes in elementary bayesian inference 169

Credible interval (95%)

Item Percent correct Lim sup Lim inf Content assessed
1 88.7 78.4 94.3 Likelihood, conditional probability
2.1 79.0 67.3 87.2 Simple probability
2.1 38.7 27.6 51.1 Conditional probability
2.2 29.0 19.2 41.2 Conditional probability of contrary
event
2.4 51.6 39.4 63.5 Joint probability
3 66.1 53.7 76.6 Comparing parameters in frequen-
tists and Bayesian inference
4 58.1 45.6 69.5 Prior distribution
5 61.3 48.8 72.3 Parameter as random variable;
statistics
6 50.0 36.6 60.4 Correct assignment of a non infor-
mative prior distribution for propor-
tion
7 93.5 84.5 97.3 Bayes’ theorem as a tool to trans-
form prior into posterior probabili-
ties
8 53.2 40.9 65.0 Parameters in Beta distribution,
defining prior informative
distribution for proportion
9 85.5 74.6 92.1 Parameters in Beta distribution
10 64.5 52.0 75.2 Computing credible intervals for
proportion; from Beta tables
11 58.1 45.6 69.5 Testing simple hypotheses for
proportion; from Beta tables
12 53.2 40.9 65.0 Properties of credible intervals
13 69.4 57.0 79.3 Posterior distribution of mean; non
informative prior; known variance
14 30.6 20.6 42.9 Testing simple hypotheses for means
15 40.3 29.0 52.7 Posterior distribution of mean; non
informative prior; unknown variance
16 69.4 57.0 79.3 Credible intervals for means
17 69.4 57.0 79.3 Posterior distribution for mean, in-
formative prior
18.1 85.8 75.4 91.3 Identifying prior probabilities from
a problem statement
18.2 74.4 63.9 83.0 Identifying likelihood from a prob-
lem statement
18.3 79.0 67.3 87.2 Bayes’ theorem as a tool to trans-
form prior into posterior probabili-
ties
19 58.1 45.6 69.5 Meaning of likelihood
20.1 82.3 70.9 89.7 Parameters in Beta curve. Spread
20.2 72.6 58.2 80.0 Parameters in Beta curve. Centre
Table 2. Percent of correct responses and contents assessed in the BLI items
170 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

of the teaching experience. A reliability analysis of responses gave a value for

the internal consistency coefficient α = 0.68, which is reasonable, given the
variety of contents (the test is multidimensional). The Pearson correlation
coefficients between BIL and final score in the statistics course ρ = 0.41: was
significant at .01 level, although moderate.
The easiest tasks were those related to distinguishing prior probabilities,
posterior probabilities and likelihood and identifying them from a problem
statement (items 7 and 1). Correct assignment of an informative distribution
for proportions (item 9), interpreting parameters in the Beta curve (items 9,
20.1 and 20.2), computing credible intervals for the mean of a normal distribu-
tion with known variance (item 16), distinguishing statistics and parameters in
a problem statement (item 17), getting a posterior distribution for the mean in
a normal population from uniform prior distribution (item 13), understanding
parameters as random variables (item 3 and 5), computing credible intervals
for proportions (item 10) were all relatively easy tasks with over 60% correct
responses on average.
There were only 4 difficult tasks (mean percentage of correct responses
under 50%). These tasks were item 14 (testing hypotheses about the mean),
where students either made a mistake in the reasoning by contradiction (choos-
ing distractor c) or did not understand the standardization operation and
choose distractor a. Of course this is a highly complex item, where the logic
of testing hypothesis is mixed with knowledge of probability calculus and
standard Normal distribution. Moreover understanding proof the logic of sta-
tistical tests and reasoning by contradiction was also difficult in other research
related to frequentist statistics (e.g. [35]).
Students also found items 2b and 2c difficult, where they confused a condi-
tional probability and its inverse, a problem that has been repeatedly reported
(e.g. in [13,32]). In comparing these results with those in item 1 and 18 where
we found a high percentage of correct responses, we remark that distractors
in item 2 are given only by formulas (instead of using a verbal expression).
We conclude that the expressions prior and posterior probabilities and
likelihood helped students to better distinguish a conditional probability and
its inverse in these items and students possibly did not remember the symbolic
expression for a conditional probability. Finding a posterior distribution for
the mean (item 15) was difficult because students forgot to divide by the
square root of the sample size to find the standard deviation in the posterior
distribution. All the other tasks had a medium difficulty (between 50–60%
correct responses).
To study the interrelations and implications between learning objectives
we carried out several multivariate analyses, using the CHIC software, Clas-
sification Hierarchical, Implicative and Cohesive [11]. The implication index
between two dichotomous variables a and b in a population is defined by (1).
 Implications between learning outcomes in elementary bayesian inference 171
 
card A ∩ B − card(A)·ncard(B)


q(a, b̄) =  q  (1)
card(A)·card(B)
n

Here A and B are the population subgroups where a and b take the value
1 [19, 33]. This index follows the normal distribution N(0, 1), and from there
an intensity for the implication a ⇒ b is defined by (2).



ϕ(a, b̄) = Pr card X ∩ Y ≤ card A ∩ B (2)

In (2) X and Y are dichotomous independent random variables having the

same cardinal as A and B respectively [27,28]. In our study we have a total of
C26.2 implication indexes among the 26 sub items in the BIL questionnaire.
The software CHIC computes these indexes and provides a graph with all the
implications which are significant to a given significance level.
The implication a ⇒ b in our study is interpreted in the sense that when
a student correctly solves item a there is higher probability for him /her to
solve item b. In this sense the implicative graph provides a possible order
to introduce the different concepts and procedures whose understanding is
assessed in the items in the teaching of the topic. Before carrying out the
implicative analysis we checked the assumptions of the method; experimental
units of variables and independence of responses by different students. We
assumed a binomial model for the responses; that is, we assumed each student
to have the same likelihood to correctly solve the items [27], as in fact these
are the hypotheses assumed in classical theory of tests [30] that was the model
used in building the questionnaire.
In Figure 1 we present the implicative graph with all the relationships
that were significant at 99% (dashed line) or 95% level (continuous line). We
observe that the implication relationship is asymmetrical and the direction of
implication is showed by the arrows in the graph.

Fig. 1. Implicative graph with significant implications at 99% and 95%

If we study the relationships higher than 99% in the graph (discontinuous

line in the diagram), we observe that students who correctly answer item 18.2
172 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

(correct identification of likelihood, which is given by a conditional probabil-

ity) have better likelihood to answer item 18.1 (correct identification of prior
probabilities, which are given by simple probabilities). Correct performance in
item 10 (computing credible intervals for the proportion, identifying probabili-
ties and critical values from the Beta distribution table and computing credible
intervals for a proportion) increases the possibility of a correct computation of
posterior probabilities with Bayes theorem (item 18.3). Both tasks involve un-
derstanding probability axioms and computing probability, although the first
one is more complex. Finally correct computation of conditional probabilities
implies correct computation of joint and single probabilities (items 2.1, 2.2,
and 2.4).
As regards implications higher than 95% (continuous line in the diagram)
we observe that students who correctly perform a Bayesian hypothesis test
(items 14 or 11) increase their likelihood to correctly interpret credible inter-
vals (item 12). In fact all the ideas and computations involved in solving the
second task are involved in the first one, which adds the need to understand
the logic of comparing probabilities for different hypotheses and in the case of
item 14 proof by contradiction. Item 14 implies item 2.3, the computation of
conditional probability for a contrary event, but, again mastering the idea
of proof by contradiction involves correct reasoning on both conditional rea-
soning and complementation. Students who visualize parameters as random
variables (item 3), or compute probabilities for Beta function and credible
intervals for proportions (item 10), perform better in correctly assigning a
Beta informative prior distribution (item 8), a task that is also facilitated by
item 14.
Item 2.3 (computing the conditional probability for the contrary event) or
item 2.2 (computing conditional probability) facilitate item 1, distinguishing
prior and posterior probabilities and likelihood (all these ideas are supported
on correct conditional reasoning). Item 2.2 facilitates computing simple prob-
ability (item 1) and both of them together facilitate the computation of joint
probabilities (item 2.4), another task which is easier for those who succeeded
in Item 14 (testing hypotheses), possibly because the idea of conditional prob-
ability is required to test a hypothesis.

5 Implicative hierarchy of learning outcomes

Once the isolate implications between items were studied we carried out an im-
plicative classification analysis to clarify the structure of implication analysed
that point to the three groups of concepts in our a-priori analysis, but that
in some points is somewhat mixed. This is an algorithm, which uses the im-
plicative indexes in a set of variables to study the internal cohesion of some
√ [12, 24]. The cohesion between two variables a y b is defined
variables subsets
by c(a, b) = 1 + H 2 where H is the entropy for the two variables, and varies
between 0 and 1. The cohesion for a class of variables [18] is defined by in (3).
 Implications between learning outcomes in elementary bayesian inference 173
 2
 r(r−1)
i∈{1,...,r−1}
Y
C(A) =  c(ai , aj ) (3)
j∈{2,...,r};j>i

Finally, given two sets of variables A and B the strength of implication

from A to B is defined [10] by (4).

" #rs

1/2
Ψ (A, B) = sup ϕ ai , bj · [C (A) · C (B)] (4)
i∈{1,...,r};j∈{1,...,s}

The software CHIC builds an implicative hierarchy in the set of variables,

taking into account the maximal cohesion into each class and the higher im-
plication from one class to another. In Figure 2 we present the hierarchy
produced.

Fig. 2. Implicative hierarchy with 95% node

There are four significant clusters:

1. Group 1: Conditional probability. Items 2.2 and 2.1 which are linked to
item 2.4, all of them related to probability. The student who correctly
computes conditional probabilities (item 2.2), correctly performs simple
(item 2.1) and compound probability (item 2.4). The higher difficulty of
conditional probability as regards simple and compound is then confirmed
as well as our previous hypothesis that conditional probability is one core
concept in the learning of Bayesian statistics.
2. Group 2: Prior and posterior distributions and Beta curves. Items 9, 7,
10, 17, 8 and the two parts of item 20. Students who are able to interpret
174 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

the parameters in the Beta curve (item 9) and understand how posterior
distributions are achieved from prior distributions and likelihood through
Bayes theorem (item 7) succeeded better in getting a credible interval
for proportions in the continuous case, a task that requires interpreting
probabilities of Beta curves, and understanding the concept of posterior
probability, as well as the concept of credible interval. They also performed
better in discriminating prior and posterior distribution of the mean (item
17). All of this leads to better choosing a non informative prior distribution
for proportion in the continuous case through the Beta Curve (item 8),
and graphically interpreting the parameters in Beta curves (item 20); both
tasks are related to understanding the meaning of these parameters. These
are a subgroup of the tasks we included in the second group of concepts
(parameters, their distribution, prior and posterior distribution) in the a-
priori analysis; specifically most of these tasks are related to Beta curves
that was a concept new to the students.
3. Group 3 (items 11, 12, 14 and 16) is a set of the concepts we included in
the third group in the a-priori analysis: Logic of Bayesian inference. Being
able to correctly test a hypothesis for proportions (item 11) increases the
likelihood of correctly interpreting credible intervals (item 12); and these
two tasks are linked to another group: correctly testing a hypothesis about
the mean (item 14), which, in turn increases the likelihood of correctly
computing a credible interval for the mean (item 16). All this knowledge
is specifically related to the logic of Bayesian methods; understanding
the test of hypotheses facilitates that of credible intervals; inference for
proportion was easier than inference for mean, possibly because students
have to distinguish in the last task the formulas for known or unknown
variance.
4. Group 4: Finally there is a second group of tasks related to conditional
probability (the different parts of Item 18, 2.3 and 1). Correct identifica-
tion of prior probability (item 18.1) facilitates the correct identification of
likelihood from a problem statement (item 18.2) and this leads to correct
computation of posterior probabilities (item 18.3). These three abilities
lead to better identification of conditional probabilities for the contrary
event (item 2.3) and discrimination between prior probability, likelihood
and posterior probabilities in the context of a problem (item1). The sep-
aration between groups 1 and 4 is explained by the different difficulty of
the tasks in the two groups. Tasks in group 4 were easier than those in
group 1 where probabilities are only given by formulas.
Other groupings of items that were non significant at the 95% level were as
follows:
1. Group 5 : Items 6 (assigning adequate prior distribution for the non in-
formative case to proportions in the discrete case), 3 (understanding pa-
rameters as random variables) and 5 (discrimination between parameters
 Implications between learning outcomes in elementary bayesian inference 175

and statistics); all these tasks are related to understanding parameters

from a Bayesian point of view, that is difficult for some students [23].
2. Group 6 : Items 13 (Posterior distribution of mean when variance is known)
and 15 (posterior distribution of mean when variance is unknown) related
to specific factual knowledge; so possibly there was no difference between
students with good or poor understanding of other concepts.
3. Group 7 : Item 4 (concept of prior distribution) and item 9 (concept of like-
lihood); both were easy items, and therefore were unrelated to knowledge
of other concepts.
In summary these implications support our a-priori analysis and point to
three groups of concepts relevant for students’ introduction to the elementary
ideas of Bayesian inference and that should be taken into account in planning
the teaching, although some of the groups still split in some subgroups:
1. Conditional probabilistic reasoning (as shown in groups 1 and 4), a theme
where many biases have been described in the literature, but which is
basic in defining posterior probabilities and likelihood, as well as in un-
derstanding the logic of credible intervals and hypothesis testing. Results
also suggested that formulas for different types of probability were harder
than verbal expressions for students to understand. Perhaps we should
take into account Feller’s suggestion ([15] p. 114) that “conditional prob-
ability is a basic tool of probability theory, and it is unfortunate that its
great simplicity is somewhat obscured by a singularly clumsy terminology”.
2. Probability distributions, its parameters (visualized as random variables),
the distinction of prior and posterior distribution of parameters and as-
signment of prior distributions for informative and non informative cases
(Groups 2, 5, 6 and 7). In our teaching we limited to Beta and Normal
distributions, since the time available for teaching was restricted, but still
so, the understanding of Beta curves appeared as a separated subgroup,
as well as remembering the rules for known and unknown variance in in-
ference about normal distributions. The difficulties to understand the dif-
ferent conception of parameters in Bayesian and frequentists statistics [23]
also appeared as a separated subgroup.
3. Logic of Bayesian inference (Group 3), that is, understanding the logic
for computing and interpreting credible intervals and testing simple hy-
pothesis. Performance in these tasks is in fact supported in understanding
the previous two groups of concepts, most of which are not specific to
Bayesian reasoning. However, limitation of teaching time leads some lec-
turers to reduce the teaching of the same and to try to pass directly from
data analysis to inference. Teaching of Bayesian inference, therefore should
only be started when previous groups of concepts are well understood by
students.
176 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

6 Discussion

In this paper we reported results from assessing students’ understanding of

elementary Bayesian ideas after a short teaching experiment. The high per-
centage of correct responses in the questionnaire (even in highly complex
tasks, such as computing credible intervals and carrying out hypothesis tests)
supports the claims for complementing the teaching of frequentists statis-
tics with some ideas of Bayesian statistics in undergraduate statistics courses
(e.g. [1, 25]). Both approaches to inference should be, however based in the
teaching of core ideas of probability and conditional probability.
A comparative analysis of the undergraduate teaching of statistics shows a
clear imbalance between what it is taught and what it is later needed; in partic-
ular, most statistics introductory courses are exclusively frequentist and many
students never get a chance to learn some Bayesian concepts which would im-
prove their professional skills ([6]). Our research shows that undergraduate
students are able to acquire an intuitive understanding for a number of con-
cepts related in elementary Bayesian inference in a short period of teaching.
The implicative and cohesive classification analyses supported our a-priori
analysis of the concepts related to understanding basic Bayesian inference and
suggested that possible difficulties are not just related to the understanding
of conditional probability. Even when the difficulties in distinguishing a con-
ditional probability and its inverse that have been repeatedly pointed out in
the literature [4,13,14] also arose in our students, its influence in their general
performance was not so high and moreover the difficulty decreased when tasks
included verbal expressions of these probabilities instead of formulas.
However, the study also provides arguments to reinforce the study of con-
ditional probability in the teaching of data analysis to psychologists, not only
because of the usefulness of this topic in clinical diagnosis, but as a base for fu-
ture study of Bayesian inference. This and other concepts that students should
have previously mastered (difference between statistics and parameters, use
of distribution tables, or operating with standard scores and inequalities) also
affected success in some of the tasks and more attention should be paid in
introductory statistics courses. At the same time, the classes obtained in the
implicative hierarchy provide us with information about the concepts whose
understanding is related and their relative difficulty. This is a potential help
to prepare didactic materials and to organize the teaching of the topic.
We are conscious this research should continue with new samples of stu-
dents. However we think we have provided arguments to introduce basic
Bayesian statistics in undergraduate courses, whenever we emphasize the ele-
ments of statistical thinking; incorporate more data and concepts, and fewer
recipes and derivations in the classroom, provide students with automate com-
putations and graphics and foster active learning [8].
 Implications between learning outcomes in elementary bayesian inference 177

Acknowledgement: This research was supported by the project SEJ2004–

00789 and grant AP2003–5130, MEC, Madrid, and FQM–126, Junta de
Andalucía, Spain.

References
1. J. Albert. Teaching introductory statistics from a bayesian perspective. In
B. Philips, editor, Proceedings of the Sixth International Conference on Teaching
Statistics, CD-ROM, 2002.
2. J.H. Albert and A. Rossman. Workshop Statistics. Discovery with Data. A
Bayesian Approach. Key College Publishing, 2001.
3. Lecoutre B., Lecoutre M.P., and Poitevineau J. Uses, abuses and misuses of
significance tests in the scientific community: Won’t the bayesian choice be un-
avoidable? ISR, pages 399–418, 2001.
4. M. Bar-Hillel. Decision Making Under Uncertainty, chapter The Base Rate Fal-
lacy Controversy, pages 39–61. North Holland, Amsterdam, 1987.
5. C. Batanero and E. Sánchez. Exploring Probability in School: Challenges for
Teaching and Learning, chapter What is the Nature of High School Student’s
Conceptions and Misconceptions about Probability?, pages 260–289. Springer,
New York, 2005.
6. J.M. Bernardo. A bayesian mathematical statistics primer. In A. Rossman
and B. Chance, editors, Proceedings of the Seventh International Conference
on Teaching Statistics. International Association for Statistical Education, CD-
ROM, 2006.
7. D.A. Berry. Basic Statistics: A Bayesian Perspective. Belmont, 1995.
8. W.M. Boldstad. Teaching bayesian statistics to undergraduates: Who, what,
where, when, why, and how. In B. Phillips, editor, Proceedings of the Sixth
International Conference on Teaching of Statistics, CD-ROM, 2002.
9. W. Bolstad. Introduction fo Bayesian Statistics. Wiley, 2004.
10. R. Couturier. Subjects categories contribution in the implicative and the simi-
larity analysis. LMSET, pages 369–376, 2001.
11. R. Couturier and R. Gras. Chic: Traitement de données avec l’analyse implica-
tive. In S. Pinson and N. Vincent, editors, Journées Extraction et Gestion des
Connaissances (EGC’2005), pages 679–684 (Vol. 2), 2005.
12. R. Couturier, R. Gras, and F. Guillet. Classification, Clustering, and Data Min-
ing Applications, chapter Reducing the Number of Variables Using Implicative
Analysis, pages 277–285. Springer-Verlag, Berlin, 2004.
13. R. Falk. Conditional probabilities: insights and difficulties. In R. Davidson
and Swift J., editors, Proceedings of the Second International Conference on
Teaching Statistics, pages 292–297, 1986.
14. R. Falk. Studies in mathematics education, chapter Inference Under Uncertainty
via Conditional Probability, pages 175–184 (Vol. 7). UNESCO, Paris, 1989.
15. W. Feller. An Introduction to Probability Theory and its Applications, Vol. 1.
Wiley, 1968.
16. J. D. Godino. Un enfoque ontológico y semiótico de la cognición matemática.
RDM, pages 237–284, 2002.
178 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

17. R. Gras. Panorama du développement de l’a.s.i. à travers des situations fon-

datrices. In R. Gras, F. Spagnolo, and J. David, editors, Troisième Rencontre
Internationale A.S.I. Analyse Statistique Implicative. Quaderni di Ricerca In
Didattica of G.R.I.M., pages 6–24 (Supplemento 2(15)), 2005.
18. R. Gras, Kuntz P., and H. Briand. Les fondements de l’analyse statistique
implicative et quelques prolongements pour la fouille de données. MSH, pages
9–29, 2001.
19. R. Gras and H. Ratsima-Rajohn. L’implication statistique, une nouvelle méth-
ode d’analyse de données. RO, pages 217–232, 1996.
20. R. Gras and A. Totohasina. Chronologie et causalité, conceptions sources
d’obstacles epistémologiques à la notion de probabilité conditionnelle. RDM,
pages 49–95, 1995.
21. L.L. Harlow, S.A. Mulaik, and J.H. Steiger. What if there were no significance
tests? Erlbaum, 1997.
22. P. Iglesias, J. Leiter, M. Mendoza, V. Salinas, and H. Varela. Mesa redonda
sobre enseñanza de la estadística bayesiana. RSCE, pages 105–120, 2000.
23. G.R. Iversen. Student perceptions of bayesian statistics. In J. Pereira-Mendoza,
editor, Proceedings of the Fifth International Conference on Teaching Statistics,
pages 234–240, 1998.
24. D. Lahanier-Reuter. Un algorithme de regroupements de modalités de variables
en analyse implicative des données. MSH, pages 5–8, 2001.
25. B. Lecoutre. Beyond the significance test controversy: Prime time for bayes?
In A. Rossman and B. Chance, editors, Bulletin of the International Statistical
Institute: Proceedings of the Fifty-second Session of the International Statistical
Institute, pages 205–208 (Tome 58, Book 2), 1999.
26. B. Lecoutre. Training students and researchers in bayesian methods for experi-
mental data analysis. JDS, pages 217–232, 2006.
27. I.C. Lerman. Classification et analyse ordinale des données. Dunod, 1981.
28. I.C. Lerman, R. Gras, and H. Rostam. Elaboration d’un indice d’implication
pour données binaires i. MSH, pages 5–35, 1981.
29. D.S. Moore. Advances in Statistical Decision Theory, chapter Bayes for Begin-
ners? Some Pedagogical Questions, pages 3–17. Birkhäuser, Stuttgart, 1997.
30. J. Muñiz. Teoría Clásica de los Tests. Pirámide, 1994.
31. A.M. Ojeda. Dificultades del alumnado respecto a la probabilidad condicional.
UNO, pages 37–55, 1995.
32. A. Pollatsek, A.D. Well, C. Konold, and P. Hardiman. Understanding condi-
tional probabilities. OBHDP, pages 255–269, 1987.
33. Gras R. L’Implication Statistique Nouvelle Methode Exploratoire de Donneés.
La Penseé Sauvage, Grenoble, 1996.
34. A. Totohasina. Methode Implicative en Analyse de Données et Application à
l’Analyse de Conceptions d’Étudiants sur la Notion de Probabilité Condition-
nelle. Ph.D. Thesis. Universidad de Rennes I, 1992.
35. A. Vallecillos. Some empirical evidence on learning difficulties about testing
hypotheses. In A. Rossman and B. Chance, editors, Bulletin of the International
Statistical Institute: Proceedings of the Fifty-Second Session of the International
Statistical Institute, pages 201–204 (Tome 58, Book 2), 1999.
 Implications between learning outcomes in elementary bayesian inference 179

Appendix: Questionnaire5

Item 1. 10 out of every 100 students in a Faculty study mathematics; 30 out

of every 100 students doing mathematics share an apartment with other
students. Let S be the event “sharing the apartment” and M the event
the student is doing mathematics course. If we pick a student at random
and the student is doing mathematics, the probability that he shares the
apartment is:
1. A prior probability P (S)
2. A posterior probability P (S|M )
3. A likelihood P (M | S)
4. A joint probability P (M ∩ S)
Item 2. Imagine you pick 1000 people at random. You know that 10 out
of every 1000 people get depression. A depression test is positive for 99
out of every 100 depressed people as well as for 2 out of every 100 non
depressed people. Given that D means depression and + means a positive
test, compute the following probabilities:
1. P (D) =
2. P (+ | D) =
3. P (− | D) =
4. P (D ∩ +) =
Item 3. The mean value for a variable (for example height) in a population:
1. Is a constant in Bayesian inference
2. Is a random variable in classical inference
3. Is a random variable in Bayesian inference
4. Could be constant or variable, depending on the population
Item 4. The prior probability distribution for a parameter:
1. Provides all the information about the population before col-
lecting the data
2. Is computed from the posterior distribution by using the Bayes theo-
rem.
3. It can be used to compute the credible interval for the parameter
4. Is an uniform distribution
Item 5. 1000 young Spanish people were interviewed in a survey. On average
they spent 3 hours a week in practicing some sports. In Bayesian inference:
1. 3 hours is a parameter in the population of young Spanish people
2. The average in this population is a random variable; the most
likely value is about 3 hours
3. The average in this population is an unknown constant
4. Each young Spanish person spends 3 hours a week in doing some sport

5
Correct responses are emphasized in bold.
180 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

Item 6. In a factory lamps are sold in boxes of four lamps. We have no

information about the proportion of defective lamps. Which of the dis-
tributions A, B. C or D better describes the prior distribution for the
proportion of defective lamps in a box?
(A) (B)
Values of Probability Values of Probability
Proportion Proportion
0.00 0.1 0.00 0.2
0.25 0.1 0.25 0.2
0.50 0.1 0.50 0.2
0.75 0.1 0.75 0.2
1 0.1 1 0.2

(C) (D)
Values of Probability Values of Probability
Proportion Proportion
0.00 0.00 0.00 1/4
0.01 0.25 0.25 1/4
0.02 0.50 0.50 1/4
0.03 0.75 0.75 1/4
0.04 1 1 1/4

Item 7. In trying to estimate a proportion a student filled three columns in

the Bayes table. He got these data:
Values of proportion Prior Probability Likelihood —— ——
0.0000 0.0000 0.0000
0.1000 0.1000 0.0000
0.2000 0.1000 0.0233
0.3000 0.1000 0.1239
0.4000 0.1000 0.0682
0.5000 0.1000 0.0065
0.6000 0.1000 0.0001
0.7000 0.1000 0.0000
0.8000 0.1000 0.0000
0.9000 0.1000 0.0000
1.0000 0.1000 0.0000
Sum 0.0222
The posterior probability that the true value of proportion in the popu-
lation is 0.4 would be:
1. 0.00682
2. 0.1000
3. 0.3072
4. 0.00015
Implications between learning outcomes in elementary bayesian inference 181

Item 8. A clinical survey showed a 15

1. B(15, 100)
2. B(15, 85)
3. B(85, 15)
4. B(100, 15)
Item 9. The mean for a Beta B(a, b) distribution is:
1. a/b
2. (a + 1)/(a + b)
3. (a + 1)/(b + 1)
4. a/(a + b)
Item 10. In the following table probabilities and critical values for the
B(30, 40) distribution are given

Probabilities Critical values

p0 P (0 < p < p0 ) P (p0 < p < 1) P (0 < p < p0 ) p0
0 0.000 1.000 0.000 0.000
0.05 0.000 1.000 0.005 0.296
0.1 0.000 1.000 0.010 0.304
0.15 0.000 1.000 0.015 0.311
0.2 0.000 1.000 0.020 0.316
0.25 0.001 0.999 0.025 0.320
0.3 0.012 0.988 0.030 0.324
0.35 0.090 0.910 0.035 0.327
0.4 0.318 0.682 0.040 0.330
0.45 0.645 0.355 0.045 0.330
0.5 0.886 0.114 0.050 0.333
0.55 0.979 0.021 0.950 0.526
0.6 0.998 0.002 0.955 0.529
0.65 1.000 0.000 0.960 0.533
0.7 1.000 0.000 0.965 0.536
0.75 1.000 0.000 0.970 0.541
0.8 1.000 0.000 0.975 0.545
0.85 1.000 0.000 0.980 0.551
0.9 1.000 0.000 0.985 0.558
0.95 1.000 0.000 0.990 0.567
1 1.000 0.000 1.000 1.000

The 98% credible interval for the proportion in a population described by

a posterior distribution B(30, 40) is about:
1. (0.316 < p < 0.551)
2. (0.304 < p < 0.567)
3. (0.3 < p < 0.6)
4. (0.1 < p < 0.9)
182 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

Item 11. The posterior distribution for the proportion of voters favorable to a
political party is given by the B(30, 40) distribution. From the above data
table, the most reasonable decision is accepting the following hypothesis
for the population proportion
1. H : p < 0.25
2. H : p > 0.55
3. H : p > 0.25
4. H : p > 0.45
Item 12. For the same posterior distribution of the parameter in a population
the r% credible interval for the parameter is:
1. Wider if r increases
2. Wider if the sample size increases
3. Narrower if r increases
4. It depends on the prior distribution
Item 13. In a normal population with standard deviation σ = 5 and with no
prior information about the population mean, we pick a random sample
of 25 elements and get a sample mean x̄ = 100. The posterior distribution
of the population mean is:
1. A normal distribution N (100, 0.5)
2. A normal distribution N (0, 1)
3. A normal distribution N (100, 5)
4. A normal distribution N (100, 1)
Item 14. To test the hypothesis that the mean µ in a normal population with
standard deviation σ = 1 is larger than 5, we take a random sample of
100 elements. To follow the Bayesian method:
1. We compute the sample mean x̄ and then compute P x̄−5


0.1 < 5 ; when
this probability is very small, we accept the hypothesis.
2. We compute x̄ and then compute P x̄−5


0.1
< Z ; when Z is
the normal distribution N (0, 1); when this probability is very
small, we accept the hypothesis.
3. We compute the sample mean x̄ and then compute P x̄−5


0.1 > Z when
Z is the normal distribution N (0, 1); when this probability is very
small, we accept the hypothesis.
4. We compute the sample mean x̄ and then compute P x̄−5


0.1 > 5 when
this probability is very small, we accept the hypothesis.
Item 15. In a sample of 100 elements from a normal population we got a mean
equal to 50. If we assume a prior uniform distribution for the population
mean, the posterior distribution for the population mean is:
1. About N (50, s), where s is the sample standard estimation.
2. About N (50, s/10), where s is the sample standard estima-
tion.
3. We do not know, since we do not know the standard deviation in the
population
Implications between learning outcomes in elementary bayesian inference 183

4. About N (0, 1)
Item 16. The posterior distribution for a population mean is N (100, 15). We
also know that P (−1.96 < Z < 1.96) = 0.95, where Z is the normal
distribution N (0, 1). The 95% credible interval for the population mean
is:
1. (100 − 1.96 · 1.5, 100 + 1.96 · 1.5)
2. (100 − 1.96, 100 + 1.96)
3. (100 · 1.5 − 1.96, 100 · 1.5 + 1.96)
4. (100 − 1.96 · 15, 100 + 1.96 · 15)
Item 17. In a survey to 100 Spanish girls the following data were obtained:
Mean Standard dev.
Sample 160 10
Prior distribution 156 13
Posterior distribution 158.5 7.9
To get the credible interval for the population mean we use:
1. The normal distribution N (160, 10)
2. The normal distribution N (156, 13)
3. The normal distribution N (158.5, 7.9)
4. The normal distribution N (160, 0.5)
Item 18. 20% of boys and 10% of girls in a kindergarten are immigrant. There
are about 60% boys and 40% girls in the center. Use the following table
to compute the probability that an immigrant child taken at random is a
boy.
Events Prior probabilities Likelihoods Product Posterior probabilities

Sum 1 1

Item 19. In a geriatric center we want to estimate the proportion of residents

with cognitive impairment. 2 out of 10 residents taken at random in the
residence showed cognitive impairment. The likelihood for the parameter
p = 0.1 is 0.1937. What is the meaning of this value?
1. P (data), that is, probability of getting this sample.
2. P (data ∩ p = 0.1), that is, probability of getting the sample and that,
in addition, the population proportion is 0, 1.
3. P (p = 0.1 | data), that is, probability of a population proportion is 0.1
given the sample
4. P (data | p = 0.1), that is, given than p = 0.1, probability of
getting this sample
Item 20. Observe the following Beta curves:
1. Which of them has a greater spread?
184 Carmen Díaz, Inmaculada de la Fuente and Carmen Batanero

a=5, b=5 a=50, b=50

8
2.0

6
f(x)

f(x)

4
1.0

2
0.0

0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

x x

2. Which of them predict a greater value of proportion in the population?

a=7, b=3 a=2, b=8

3.0
2.0

2.0
f(x)

f(x)
1.0

1.0
0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

x x
Personal Geometrical Working Space:
a Didactic and Statistical Approach

Alain Kuzniak

Equipe Didirem Université Paris 7 France

[email protected]

Summary. In this paper, we study answers that pre-service teachers gave in an

exercise of Geometry. Our purpose is to gain a better understanding of what we call
the geometrical working space (espace de travail géométrique). We first conduct a
didactical study based on the notion of geometrical paradigms that leads to a clas-
sification of student’s answers. Then, we use statistical tools to precise the previous
analysis and explain students’ evolution during their training.

Key words: Geometry, Didactic, Paradigm, Geometrical Working Space, Teachers

Training.

1 Presentation of the study

Various theoretical tools have been developed to study the teaching of geome-
try and, in the case of teacher training, two of them have been here preferred:
geometrical paradigms and geometrical working spaces (GWS; in French:
Espace du Travail Géométrique). Using these tools, our research focused on
the following hypothesis, which our work abundantly supports:
In education, the sole term geometry evokes several distinct para-
digms. By and large, these paradigms reflect the breaks observed be-
tween the various academic cycles in the teaching and learning of
geometry.
In our view, the field of geometry can be mapped out according to three
paradigms, only two of which — Geometry I and II — play a part in today’s
secondary education. Each paradigm is global and coherent enough to define
and structure geometry as a discipline and to set up respective working spaces
suitable to solve a wide class of problems. Based on these premises, we built
a training device designed to make future teachers aware of these paradigms
and of their role as cause of certain misunderstandings in a classroom setting.
The construction and evaluation of the device requires a precise analysis of
A. Kuzniak: Personal Geometrical Working Space: a Didactic and Statistical Approach,
Studies in Computational Intelligence (SCI) 127, 185–202 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
186 A. Kuzniak

students’ spontaneous use of paradigms as they solve geometrical problems.

This analysis is meant to understand better the geometrical working space of
each student. Existing research provided the elements that lead to distinguish-
ing among four groups of students, each corresponding to a specific approach
to the study of geometry.
In this paper, we wish to examine what specific contribution statistical
methods can bring to that research. More precisely, we focus on the two fol-
lowing sets of questions:
• The first set bears on the classification resulting from our didactical analy-
sis. Does statistical analysis produce the same outcomes as the initial
analysis? Which new elements, if any, emerging from implicative analy-
sis, help better understand the various classes of students and thus predict
some of the changes observed during training sessions?
• The second series of questions is concerned with automating the process of
sorting students according the classes defined above. Indeed, in addition
to being demanding, the didactic analysis calls for advanced knowledge
of its theoretical framework, which limits its use by other researchers. To
mitigate this problem, we are currently working with Chilean colleagues
on developing tools that will enable them to analyze large quantities of
data on student performance.

In this paper, we first expand on our theoretical framework in some detail

and present the training device. Then we examine the device’s key exercise.
We will present data on student performance that highlight the role and con-
tribution of the methods we used: a didactic analysis and then a statistical
study (factorial and implicative analysis).

2 Object of the study

2.1 Theoretical premises

Work initiated by Bachelard [1] and Koyré [2] and pursued in mathematics
by Lakatos [3] showed that the idea of a peaceful scientific evolution of math-
ematical concepts was an illusion. Kuhn [4] brought the conflicting logic of
scientific ideas to a culmination: he sees the transition from one paradigm to
another as a revolution whereby a new paradigm replaces the old one.
Our view of the study of geometry is based on an approach asserting that
geometry has undergone significant changes of perspectives equivalent to par-
adigmatic shifts. Following Gonseth [5] who places geometry in relation to
the problem of space and applying Kuhn’s notion of a paradigm, we con-
sider three geometrical paradigms [6, 7] that organize the interplay between
intuition, deduction, and reasoning in relation to space:
Personal Geometrical Working Space: a Didactic and Statistical Approach 187

• Natural Geometry (Geometry I), which finds it validation in reality and

the sensible world. In this Geometry, an assertion is accepted as valid using
arguments based upon experiment and deduction. The confusion between
the model and reality is great and any argument is allowed to justify an
assertion and convince;
• Natural Axiomatic Geometry, whose archetype is classic Euclidean Geom-
etry. This Geometry (Geometry II) is built on a model that approaches
reality. Once the axioms are set up, proofs have to be developed within
the system of axioms to be valid;
• Formalist Axiomatic Geometry (Geometry III), in which the system of
axioms itself, disconnected from reality, is central. The system of axioms
is complete and unconcerned with any possible applications in the world.
These various paradigms — and this is originality of our approach — are
not organized into a hierarchy, one is not better than the other: their use is
different depending on the aim of the problem.
Our theoretical framework is also based on the notion of Geometrical
Working Space (GWS) (see 4.4 for some detail) which enables us to ana-
lyze how students or experts work when they are involved in a geometrical
task.

2.2 The question of the teachers’ training

We have examined students’ application of geometrical paradigms and use of

personal working space in various ways. The approach presented here relies
on a relatively complex training device applied in two different phases [8].
Students are primary school teachers in training.
The first phase is based on a written individual questionnaire. Specifi-
cally, students are asked to solve geometrical exercises and list the doubts and
difficulties they experienced during the résolution. During the second phase,
students are asked in particular to participate to a work entitled: “Geometry:
Charlotte and Marie, who is right and why? The students do not agree”. For
the purpose of this work, they look at a selection of solutions and comments
written by their peers during phase 1. The solutions and comments were
grouped in four categories that reflected the different approaches encountered
in students’ responses. Then students review their own initial answers.
In the next sections, we present the problem submitted to the students,
then we expose different methods used to analyze in depth the solutions given
by the students.

2.3 The key problem “Charlotte and Marie”

The following problem (Hachette Cinq sur Cinq 4e 1998, page 164) exemplifies
the kind of geometrical exercises for which the existence of a working space
suitable to solve the problem is not obvious.
188 A. Kuzniak

1.Why can we assert that the

quadrilateral OELM is a rhom-
bus?
2.Marie maintains that OELM is
a square. Charlotte is sure that
it is not true.
Who is right?

The drawing looks like a square but its status in the problem is not clear.
Is the drawing a real object the problem suggests to study or does it result
from a construction described in a text? And in that case, is the practical
achievement essential or does it only serve as a support for reasoning?
The function of the represented object is usually given by the text of the
problem: this in turn orients towards a precise geometrical paradigm. Here,
the wording gives no such indications and as a student points out: There are
no texts for the wording, only a drawing that can mislead.
Finally, who is right? Charlotte or Marie? Pythagoras’ theorem, which
doesn’t require the real measurement of the angle, gives a typical way of
handling this kind of exercise. But even there, the ambiguity of the choice of
the working space reappears. For our purpose, we shall introduce two forms
of Pythagoras’ theorem, the usual one, an abstracted form, with real numbers
and equalities:
If the triangle ABC is right in B then AB 2 + BC 2 = AC 2
and the other one, a practical form, using approximate numbers and, in a less
common way, approximate figures
If the triangle ABC is “almost” right in B then AB 2 + BC 2 ' AC 2
The first form leads to work in Geometry which deviates from data of
experiment by arguing in the numeric setting. The second formulation appears
rather as an advanced form of Geometry I.
If we work in Geometry II by using the abstracted form of Pythagoras’
theorem, then we can argue, as one student suggests, giving reason to Char-
lotte:
We know that if OEM is right in O then we have OE 2 +OM 2 = M E 2
We verify 42 + 42 = 5.62 and 32 6= 31.26. Thus, OEM is not a right
triangle.
If we use the practical Pythagoras’ theorem in the measured setting then
we shall rather follow the reasoning proposed by another student who con-
cludes:
√
Marie is right OELM is a square since 32 ' 5.6.
Personal Geometrical Working Space: a Didactic and Statistical Approach 189

In fact, it would be necessary to conclude that OELM is “almost” a square.

But, for lack of a suitable language, students cannot play on these various dis-
tinctions. They are faced at the same time with an epistemological and didac-
tical misunderstanding. It seems to us that the interplay between Geometry
I and Geometry II explain and work on this problem.
The problem comes from a textbook designed for 14 years old students and,
as noticed above, it is especially ambiguous. Its use in a class should be ques-
tioned within the framework of high-school Geometry teaching in France [9].
Why give it to seek with the pupils? And with which teaching intentions? In
our specific study, we gave it to pre-service teachers with the two main goals
of bringing out their knowledge in geometry and making explicit some mis-
understandings existing in the teaching of geometry with help of geometrical
paradigms.

3 Didactic analysis of students’ works

3.1 Towards a classification of the students’ answers

From the answers given by the students, we can sketch a classification that
takes into account the geometrical paradigm applied in the resolution. It must
be clear that only answers and not the students are classified here. But, by
doing this, a general understanding of students’ behavior is intended.
We have also identified four kinds of answers to the “Charlotte and Marie”
problem. This allows us to bring out four main approaches.
We labeled these four groups PII, PIprop, PIperc, and PIexp, for rea-
sons that will be clarify below. In each case, we will give a typical answer
(Appendix 1) from the sub-population under study.
First, answers using theorem are common among two groups of students,
PII and PIprop.
PII In this case, [St A], the standard Pythagoras’ theorem is applied
inside the world of abstract figures and numbers without considering the real
appearance of the object. Only information which is given by words and signals
(code of segments, indications on the dimension of the lengths), is used, and
Pythagoras’ theorem is applied in its entire formal rigor. To prove that the
quadrangle is a rhombus (four sides of the same length) and to show that it is
not a square (contrapositive of Pythagoras’ theorem), students use minimal
and sufficient properties. We shall consider this population as being inside
Geometry II.
PIprop. This population groups together students who apply the practical
Pythagoras’ Theorem, in fact, to be rigorous, the converse. They generally
conclude that Marie is right [St B]. In that case, the students recognize the
importance of the drawing and of the measurements’ approximation. The
practical Pythagoras’ theorem appears as a tool of Geometry I. We have
designated this population as PIprop to insist on the fact that individuals
190 A. Kuzniak

of this group use properties to argue. The question whether these students
can play with the differences between Geometry I and Geometry II or if their
horizon remains only technological.
An addition to these answers, here are those of students who did not use
Pythagoras’ theorem.
PIexp. We group together students who use their measuring and drawing’s
tools to arrive at an answer. They are situated in the experimental world of
Geometry I. Generally, this type of students concludes that Marie is right [St
C]. But, it is not always the case: a student, using his/her compass, verifies
that the vertices of the quadrangle are not cocyclics and s/he can assert that
OELM is not a square.
PIper. In this last category [St D], we group together students whose
answers are based on perception: Their interpretation of the drawing is the
basis for their answer, and they do not give use any information about their
tools of investigation. It is not easy to know if this lack of deductive proof is due
to a lack of geometrical knowledge or to a real confidence in the appearance of
the figure. To answer to this question, we must have a look at their reasoning
problems.

3.2 A look at reasoning problems

The typical outcomes presented above are logically quite coherent and do
not contain too many reasoning errors and formulation problems. That is not
true for all cases and we proceeded by performing an analysis of proofs and
the reasoning structure based on the levels of argumentation inspired by Van
Hiele.
We classify in level 1 works, which enumerate a non-minimal list of quad-
rangle properties to justify assertions. In level 2, we place productions, which
evoke a correct relation of inclusion between square set and rhombus set. In
level 3, we set productions that use minimal and sufficient information to
justify assertions.
This analysis allows us to separate two categories of students. In the first
one, widely illustrated by our previous examples, students have solid knowl-
edge concerning the figures’ properties and use level 3 reasoning. The students
of the second category argue with an accumulation of properties and show not
very sound knowledge of the geometrical properties. Here are two examples
illustrating this second group. [St E]
1) The quadrangle OELM is a rhombus. It follows the characteristics
of such a figure: four sides are equal; diagonals cut themselves in their
middle and form a right angle.
2) Both girls are right; OELM is a square, for it has four equal sides
and four right angles. It is also a rhombus, even if this figure that is
rhombus is not necessarily composed of right angles.
Personal Geometrical Working Space: a Didactic and Statistical Approach 191

This student justifies his/her first answer by enumerating a list of prop-

erties of rhombuses. Thus, we classify his/her production at level 1. The
properties employed are partially justified through visual or instrumented in-
dications. This student considers the figure in its material reality and her
approach of the problem comes within Geometry I.
The answer to the second question both girls are right occurs frequently
enough. Its justification shows that the statement Charlotte is sure that it
is not true is wrongly interpreted as Charlotte asserts that it is a rhombus
concealing the assertion It is not a square. The student focuses on the question
of the link between squares and rhombuses. It is a classic question (but not
asked here) and the student knows how to answer. That shows that she has
level 2 knowledge corresponding to the classification of figures.
With this student, we meet a rather frequent profile. [St F]
1) Four sides of the quadrangle are parallel between them and of the
same length OE = ML and OM = EL.
Definition of the rhombus: we can say that diagonals have the same
middle and are perpendicular between them.
2) Marie is right, OELM is also a square because sides are all of the
same length: OE = M L = EL = OM = 4cm.
Let us remember that the square is also a rhombus but which has
the peculiarity of having all sides with the same length (thus forming
right angles) and having diagonals of the same length.
The employed syntax could refer to level 3: some partially correct implications
are evoked. But the body of knowledge is not very reliable. In particular, we
find here a rather frequent pupil’s theorem: Any quadrangle having four equal
sides is a square. We are clearly within Geometry I where visual indications
are used to support reasoning.

4 The statistical analysis

4.1 Limit of the didactic study

The didactical study leaves some of the questions raised in introduction pend-
ing. The classification we obtained is a straight product of our theoretical
framework. It is therefore compelling to use statistical tools to test the model
and, at the same time, measure the distance between the classes. In phase 2 of
the exercise, students had to choose the best explanation among the variety
of responses. It turns out that the choices they made depended on the class
they belonged to, plus other unknown reasons. To further the analysis, we
need to understand what determines students’ class membership and define
the relevant sub-classes so that we can explain why members of the same class
could evolve in different ways during the teaching process.
192 A. Kuzniak

Moreover, we need to increase our capacity to analyze large amounts of

data through automation to learn more about students’ personal GWS and to
make this kind of research more accessible to researchers with different theo-
retical backgrounds. Let us note finally that if statistical techniques can put
our didactic approach to the test, the latter can in turn help verify whether
statistical tools are adequately discerning and explaining the phenomena un-
der study.
We chose two statistical approaches to handle the data; the first is based on
factorial analysis, the second uses implicative analysis. To make these different
analyses we retain only the productions of two groups of French students, that
are 57 subjects.

4.2 The factorial approach

Even if our population is small, we first used principal component analysis

to get a global view of data. We encoded and analyzed the answers given by
students to three questions set within the framework of the problem “Charlotte
and Marie” (Appendix 2).
This encoding, which was performed in association with J.C Rauscher and
Chilean colleagues of the University of Valparaiso, allows for the eight com-
ponents dimension (named here aspect) that underlie the universe of answers,
yielding 13 binary variables.
We introduced two components to describe the first question on the rhom-
bus. Aspect 1 represents the sources of information used by the student: does
s/he take information directly from the drawing or not? The various justifi-
cations the student uses to prove that the figure is a rhombus are represented
in Aspect 2: accumulation of arguments, characteristic property.
The answer given by the student to the question 2 — who is right?
Charlotte or Marie — is kept in aspect 3. Let us note that 20 students an-
swered Charlotte, 28 Marie, 7 of them both and two students asserted we could
not know. In order to deal with this aspect, we need two 0–1 variables CHA
and MAR. The first one CHA takes value 1 if the student answered Charlotte;
the second MAR is 1 when the answer is Marie. With this coding, the answer
both gives so values 1 to CHA and MAR at the same time. The arguments
given by the student are listed in aspect 4: reference to a theorem, use and
type of calculations, remarks on angles or sides, correction and coherence of
the reasoning.
The possible use of the triangle congruence cases is taken into account
by aspect 5 (suggested by the Chilean colleagues, this use never appeared
in France). Aspect 6 was introduced in order to check if students consider a
relation between rhombuses and squares to support their argumentation.
Some students add marks on the drawing submitted in the problem (aspect
7). The variable FIG takes the value 1 as soon as a mark or a construction is
made on the drawing.
Personal Geometrical Working Space: a Didactic and Statistical Approach 193

The last question about the doubts of the students is taken into account
by aspect 8. This aspect is divided in three variables depending on the content
the students mention: properties, drawing and estimation.
These eight aspects are then shaped into disjunctive (Yes/No) variables to
allow the statistical analysis: that gives 14 characters. We keep neither aspect
5 nor aspect 4c in this analysis.
The study was made with the program Statistica, we give here (fig. 1) the
representation of variables in the prime factorial plane.

Fig. 1. Factorial analysis

Expressing the answer to the problem, Charlotte (CHA) and Marie (MAR)
are obviously the most determining variables and it could have been interest-
ing to consider them as supplementary variables [10, 11]. With the statistical
study, we can correlate these two variables with the others which seem in our
view to be important, as the use of square roots (RAC) or of drawings on the
figure (FIG). Both variables CAR and ACC are pointing in the characteristic
properties use and are determining for a better understanding of students’
reasoning.
The graph shows the three variables connected to the doubts described by
the students: DES for doubts on drawing, APP for the estimation and finally
PROP for the expression of problems linked to the properties. Showing the
relation between rhombus and square, the place of variable LOS is also going
to be interesting to evaluate student’s reasoning level.
We should bear in mind that in the didactic approach the subgroup PIperc
includes the students who answered Marie but without our knowing the ex-
act nature of their reasoning: is the conclusion they give based on the sole
perception or given by default due to the lack of geometrical knowledge and
the neglect of certain properties? We introduced the analysis with levels argu-
mentation to better understand the method used by these students, due to its
194 A. Kuzniak

complexity, we were able to make only a qualitative analysis, the statistical

analysis complete the first study in a more systematic way.

4.3 The implicative approach

Thanks to factorial analysis, we can first sketch a map that positions students
in the prime factorial plane. To ascertain a better grouping of the determining
variables, we used the program CHIC and created similarity trees as well as
hierarchic trees that reveal one-way relationships among variables [12, 13].
This approach is essential to studying how students’ geometrical thinking
works and describe their GWSs in a more dynamic way [14, 15].

Fig. 2. Similarity tree

Fig. 3. Implicative tree

The implicative study produces four clearly defined, potentially interesting

groupings that we combine with the results of our didactic analysis. Variables
Personal Geometrical Working Space: a Didactic and Statistical Approach 195

[CHA, SRAC, PYT, CAR, IND] create a first set which shows us how stu-
dents giving the answer Charlotte (CHA) are reasoning. They use Pythagoras’
theorem (PYT) with a calculation without square roots (SRAC), they master
the notion of characteristic property (CAR) and use only information given
by the problem wording (IND). This group is very close to the one that we
identified under the name of PII.
Another group is organized around variables [LOS, MAR]. The implicative
analysis confirms that students having told from relations between rhombus
and square answered Marie (MAR). This group is close to the one described by
variables [ACC, PROP, FIG], but if students of this group lean their reasoning
on the figure and give a series of properties, at the same time they indicate
their doubts and their difficulties. The students from these two classes argue
differently from those belonging to the first group (PII): visual or experimental
use of the support of the figure, accumulation of arguments or reasoning based
on the global perception of the figure shape.
In a specific way, the statistical analysis shows the coherence of a last
group around [APP, RAC, COR]. These answers use the “approximate” form
of Pythagoras’ theorem. It seems that students from this group are sensitive
to the importance of the estimation and to the question of the drawing which
looks like a square. The way they argue is close to group PII but their regard
to the reality is different.
The combination of both analyses allows the organization of the variables
as shown in the graph of the binary variables in the prime factorial plane
(fig 4).

Fig. 4. Factorial analysis

4.4 Precision on the components of the geometrical working space

We must further interpret the results presented above in terms of geometrical
working space and for this purpose we first provide some additional elements
to clarify a notion so essential to our approach to geometry teaching.
196 A. Kuzniak

More precisely [16], the Geometrical Working Space (GWS) is the place
organized to ensure the geometrical work. It makes networking the three fol-
lowing components:
• the real and local space as material support,
• the artifacts as drawings tools and computers put in the service of the
geometrician,
• a theoretical system of reference possibly organized in a theoretical model
depending on the geometrical paradigm.
The geometrical working space becomes manageable only when its user
can link and master the three components above.
To solve a problem of geometry, the expert has to work with a suitable
GWS. This GWS must meet two conditions: its components are sufficiently
powerful to handle the problem into the right geometrical paradigm and,
depending on the user, its various components are mastered and used in a valid
way. In other words, when the expert has recognized the geometrical paradigm
involved in the problem, she/he can solve it thanks to the GWS suited to this
paradigm. When the problem is set a person (the pupil, the student or the
professor), either an ideal expert, this person handles the problem with its
personal GWS. This last one will have neither the wealth nor the performance
of the GWS of an expert.
This focus on the personal GWS, led us to introduce a cognitive dimension
into our GWS approach. For that purpose, we follow Duval [17, p 38] who
points out three kinds of cognitive processes:
• visualization process with regard to space representation and the material
support,
• construction process depending on the used tools (rules, compass) and on
the configuration,
• reasoning in relation to a discursive process.
These three processes are linked in a diagram (fig. 5) we juxtapose with
GWS components.

4.5 An interpretation in terms of personal Geometrical Working

Space

Now we can interpret the results of the statistical analysis of the students’
answers in connection with the notion of personal GWS. In terms of GWS,
clearly the study points out two systems of reference: one associated to Geom-
etry I and the other to Geometry II. The new dimension brought by the
statistical data is that technical mastery of reasoning and properties knowl-
edge introduces differences between personal students’ GWSs. The influence
of visualization or artifacts changes according to the student
A first students’ group works inside the GWS/GII based on Geometry II
system of reference. We can divide this set in two subgroups. Students of the
Personal Geometrical Working Space: a Didactic and Statistical Approach 197

Fig. 5. Global GWS Structure

first groups master enough -at least in the exercise Charlotte and Marie- the
theoretical system of reference. This group matches exactly to PII represented
above by the production of student A. Within the limits of this exercise, this
group masters rules of the geometrical argumentation. When members of this
group evoke doubts on the drawing, they underline the misleading aspect of
the drawing following the traditional view about figure in French geometry
education as soon as Geometry II is set up in the curriculum (Grade 7 or 8).
The second subgroup refers always to Geometry II but with an insufficient
mastery due either to the neglect of certain properties, or to the superficial
understanding of reasoning rules in Geometry II, encountered during their
studies. This subgroup is strictly included in PIperc, a group we noticed the
heterogeneity. We meet students here whose answers is similar to student D
but who express their doubts in a particularly subtle way as this one:
Could we say that diagonals are really perpendicular? Could we say
that the quadrangle has 4 right angles? By using a set square, yes. By
calculating with Pythagora, it is not exact, but approximate.

5.62 = 31.36 ' 32

The second large population that the analysis points out groups together
students who have in mind Geometry I paradigm and work into the working
space GWS/GI To them, the figure given into the problem is a real object
they have to study. The analysis reveals two subgroups, members of the first
use mainly arguments based on visualization and construction to solve the
problem, members of the second use the connection between construction
and proof. In this population, we meet vague answers close to student D
(PIperc) but also to student C who used drawing instruments to verify prop-
erties (PIexp). Part of the students who used the “approximate” Pythagoras’
theorem (PIprop) belongs to this large group.
198 A. Kuzniak

Finally, it is necessary to point out a group which seems to play the game
(GI/GII). These students look to balance between GI and GII but the usual
rules of the didactic contract leave them few possibilities of expressing clearly
their opinion. Members of this group give answers close to student B and
use like him/her the “approached” form of Pythagoras’ theorem but some can
also have used the classic Pythagoras’ theorem by writing their doubts on the
status of the figure drawn in the problem.
Chart summarizing the results:

Fig. 6. Relations between GWS and Students types

5 Conclusion

We set out to create a typology of teachers in training (pre-service) who were

given a series of geometry problems. The typology is based on students’ knowl-
edge of geometry which, in this study, was considered within the context of the
paradigms that give its various meanings to geometrical work and thinking.
We arrived at the notion that the Geometric Working Space depends on the
geometric perspective adopted — in our terminology Geometry I or II — and
on the user who adopts it. This observation led us to focus on the personal
GWS for each student. Using statistical techniques — in particular implica-
tive analysis — we were able to characterize students’ output by exposing
how students organize their reasoning within a finite number of categories.
We also gained a better understanding of why students evolve, or not, during
training.
The study suggests the existence of different categories of students. But
it also shows that we should consider these results with caution. We should
not assign students to rigid categories but rather be mindful of these classes
as useful benchmarks in the training of teachers.
Personal Geometrical Working Space: a Didactic and Statistical Approach 199

Moreover, personal GWS could depend on national learning curricula as

we observed it by comparing Geometry teaching in Chile and France. Ex-
planations and writing of results are not provided on the same way and new
arguments based on nature’s numbers (irrationality) or on congruent triangles
have been used by Chilean students. As the same paradigm could be tack-
led on different way in various countries and institutions, we need studies of
geometrical working spaces in various contexts.
Acknowledgment
This study would not have been possible without the active participation
of J.C. Rauscher IUFM d’Alsace.
On the other hand, the project enters a research with Chile supported by
ECOS/CONICYT.

References
1. G. Bachelard. La formation de l’esprit scientifique. Vrin Paris (Translation For-
mation of the Scientific Spirit (Philosophy of Science), Clinamen Press, 1983.
2. A. Koyré. From the Closed World to the Infinite Universe. (Hideyo Noguchi
Lecture), Johns Hopkins University Press; New Ed edition, 1969.
3. I. Lakatos. Proofs and Refutations: The Logic of Mathematical Discovery.
Cambridge University Press, 1976.
4. T. S. Kuhn. The Structure of Scientific Revolutions (Foundations of Unity of
Science). University of Chicago Press, 2Rev Ed edition, 1966.
5. F. Gonseth. La géométrie et le problème de l’espace. Griffon Ed, Lausanne,
1945–1952.
6. C. Houdement, A. Kuzniak. Sur un cadre conceptuel inspiré de Gonseth et
destiné à étudier l’enseignement de la géométrie en formation des maîtres. Ed-
ucational Studies in Mathematics, volume 40/3, pages 283–312, 1999.
7. C. Houdement, A. Kuzniak. Elementary geometry split into different geo-
metrical paradigms. Proceedings of CERME 3, http://www.dm.unipi.it/
~didattica/CERME3/proceedings/Groups/TG7/, 2003
8. A. Kuzniak, J. C. Rauscher. On Geometrical Thinking of Pre-Service School
Teachers. Cerme IV Sant Feliu de Guíxols Espagne, http://cerme4.crm.es/
Papers\%20definitius/7/kuzrau.pdf, 2005.
9. R. Berthelot, M. H. Salin. L’enseignement de la géométrie au début du collège.
petit x, volume 56, pages 5–34, 2001.
10. P. Orus, P. Gregori. Des variables supplémentaires et des élèves fictifs dans la
fouille des données avec CHIC. Actes des troisièmes rencontres ASI Palerme,
pages 279–292, 2005.
11. A. Scimone, F. Spagnolo. The importance of supplementary variables in a case
of an educational research. Actes des troisièmes rencontres ASI Palerme, pages
317–326, 2005.
12. R. Gras. L’implication statistique: nouvelle méthode exploratoire de données.
La pensée sauvage, 1996.
13. P. Kuntz. Classification hiérarchique orientée en ASI. Actes des troisièmes ren-
contres ASI Palerme, pages 53–62, 2005.
200 A. Kuzniak

14. S. Ag Almouloud. Une étude diagnostique en vue de la formation des en-

seignants. Annales de Didactique et de Sciences Cognitives., Volume 9, pages
223–246, 2004.
15. M. Bailleul, Ratsimba-Rajohn. Analyse de la gestion des phénomènes
d’ostension et de contradiction par l’analyse implicative. In Colloque Méthodes
d’analyses statistiques multidimensionnelles en didactique des mathématiques,
Caen, pages 199–216, 1995.
16. A. Kuzniak. Paradigmes et espaces de travail géométriques. Canadian Journal
of Science and Mathematics. volume 6.2, pages 167–187, 2006.
17. R. Duval. Geometry from a cognitive point of view in Mammana Perspectives
on the Teaching of Geometry for the 21st Century: An ICMI Study, Kluwer,
pages 37–51, 1998.

Appendix 1 Student’s answers

PII [Student A]
1) OELM is a rhombus because its successive sides are equal.
2) If OELM is a square, then MEL is a right-angled triangle L. Ac-
cording to the Pythagoras’ theorem we would have then, M E 2 =
M L2 + LE 2 as M L2 + LE 2 = 16 + 16 = 32 and M E 2 = 5, 62 = 31.36
Thus, angle ELM is not a right angle.
Consequently, OELM is not a square and it is Charlotte who is right.
PIprop [Student B]
1) OELM is a rhombus, for OE = OM = M L = LE and a rhombus
has its four sides of the same length.
2) Marie is right because all the sides of the quadrangle have the
same length and there is at least a right angle. We can verify it by
Pythagoras’ theorem. M E 2 = M L2 + LE 2
42 + 42 √= 16 + 16 = 32
M E = 32 = 42 ' 5.6 thus M LE = 90
PIexp [Student C]
1) OELM is a rhombus, for its diagonals cut themselves in their middle
(measuring) by forming right angles (using a set square).
Remark: the student built the second diagonal on the figure.
2) Marie is right. It is a square, for besides being a rhombus, OELM
has its angles right (set square).
PIperc [Student D]
1) Four sides of the quadrangle are parallel between them and of the
same length OE = M L and OM = EL. According to the definition
of a rhombus, we can say that diagonals have the same middle point
and are perpendicular between them.
2) Marie is right; OELM is also square because its sides form a right
angle.
Personal Geometrical Working Space: a Didactic and Statistical Approach 201

Appendix 2 Codification of the problem Charlotte

and Marie

Question 1 Why is it a rhombus?

Aspect 1: source of the information IND
Code 1 Exclusive use of the information given in the wording.
Code 2 Use of information not given explicitly by the wording.
Aspect 2: Why is it a rhombus? CAR and ACC
Code 1 correct Justification based on a necessary and sufficient condition
using the sides.
Code 2 correct Justification based on a necessary and sufficient condition
but without using sides.
Code 3 Justification using an accumulation of arguments.
Code 4 Wrong justification
Question 2 Who is right? Why?
Aspect 3: Who is right? CHA and MAR
Code 1: Charlotte
Code 2: Marie
Code 3: we cannot know
Code 4: both
Aspect 4a) Justification of the question 2 Why? PYT
Code 1: Refer to the Pythagoras’ theorem
Code 2: does not make reference to the Pythagoras’ theorem
Aspect 4b) (Calculations) RAC and SRAC
Code 1: calculations without square roots
Code 2: calculations with square roots
Code 3: without calculations
Aspect 4c) (Arguments)
Code 1: only with angles
Code 2: only with sides
Code 3: with angles and sides
Code 4: with the diagonal
Aspect 4d) COR
Code 1: relevant Arguments
Code 2: wrong Arguments
Other remarks
Aspect 5: Use of the congruence of triangles
Code 1: yes
Code 2: no
Aspect 6: relation between squares and rhombuses LOS
Code 1: reference to the relation between squares and rhombuses
Code 2: no reference to this relation
Aspect 7: Presence of marks or lines on the drawing FIG
Code 1: visible tracks or reference to use of instruments
Code 2: no visible Tracks
202 A. Kuzniak

Question on the doubts and difficulties

Aspect 8a) (Knowledge) PROP
Code 1: doubts on the knowledge of the definitions and of the properties
Code 2: no doubts expressed on this point
Aspect 8b) (Drawing)
Code 1: doubts about the status or the information given by the drawing
Code 2: no doubts on this point
Aspect 8c) (Estimate) APP
Code 1: doubts on the estimate
Code 2: no doubts on this point
 Part III

A methodological answer in various application

frameworks
Statistical Implicative Analysis of DNA
microarrays

Gerard Ramstein

LINA, Polytech’Nantes
Rue Christian Pauc BP 50609 44306 Nantes cedex 3, France
[email protected]

Summary. This chapter presents an application of the Statistical Implicative

Analysis to microarray gene expression data. The specificity of these data requires an
adaptation of the concept of intensity of implication. More specifically, we propose
to study the rankings of observations instead of the measurements themselves. This
method makes our analysis more robust and insensitive to any monotonic transfor-
mation of gene expression. We introduce the concept of rank interval and show that
the integration of the implicative method in this framework is more efficient than
correlation techniques. Our method is applied to the most challenging problems en-
countered in gene expression analysis, namely the discovery of gene coregulation,
gene selection and tumour classification. We compare our method with performing
algorithms that are dedicated to gene expression data or that are well-suited to
high-dimensional variable space.

Key words: ranking analysis, feature selection, classification rules, gene coregula-
tion, microarray data analysis

1 Introduction

The microarray (DNA chip) technology [23] monitors a consequent subset of

the whole genome on a single chip, so that biologists can observe the inter-
actions among thousands of genes simultaneously. This powerful technique
revolutionizes the traditional methods in molecular biology generally consist-
ing in experiments focused on one well-defined gene. Gene expression analysis
tends to be a major issue in data mining [21], involving new prospects in drug
discovery and disease diagnosis. From a theoretical point of view, microarray
analysis provides a deeper insight into the processes involved in the mystery
of life. This paper introduces the application of Statistical Implicative Analy-
sis [15] to these challenging data. Our objective is to extract hidden structures
from noised data, in the context of unsupervised and supervised analyses.
For unsupervised analysis, the discovery process makes no assumption on the
observations. For supervised analysis, it uses the class information relative
G. Ramstein: Statistical Implicative Analysis of DNA microarrays, Studies in Computational
Intelligence (SCI) 127, 205–225 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
206 Gerard Ramstein

to experimental conditions. The latter corresponds to microarray studies in

which the phenotypes (i.e. diseases) are known. Linking clinical phenotypes
to genotypes is a major issue for tumour classification and drug discovery.
DNA microarrays may be used to discriminate different tumour subtypes by
monitoring gene expression profile on a genomic scale.
Owing to the imprecision of the measurements, we consider the relative ranks
of the observations instead of their values. In the analysis of microarrays, the
study of the ranks presents several advantages. As it is scale-independent,
it is not sensible to the natural variation of gene expression. On the other
hand, ranking is a useful indicator for the biologist. High ranking (resp. low
ranking) corresponds to observations presenting an over-expresssion (resp.
under-expression). These expression levels play a crucial role for the analysis
of microarray data. The definition of these levels is however very difficult: it is
not possible to define absolute thresholds of expression representing a typical
expression level. Our rank analysis overcomes this difficulty as it only con-
siders the order of the data without any partitioning. The paper is organized
as follows. Section 2 surveys related work in the area. Section 3 introduces
some concepts and defines our implicative measure. We notably show that
it is more efficient than correlation measures for the analysis of gene expres-
sion profiles. In section 4 we address the problem of tumour classification. We
apply our method to extract the most informative genes that discriminate
different tumour subtypes. We use the implicative technique for the discovery
of classification rules and compare our results with different techniques of tu-
mour classification. Finally, section 5 presents an analysis of gene association
network, based on the concept of intensity of implication.

2 Related work

Dealing with imprecise and noisy data is an important issue that has already
been addressed by the researchers in the area of implicative analysis. Their
works put emphasis on the determination of intervals. In [14], an optimal
partition on numeric variables is defined and the quality of implication is
determined by the union of elements of the partition. Another interesting
work [13] introduces fuzzy partitions. We do not use either of these approaches
because we prefer avoiding a partioning procedure. We indeed observed that
microarray datasets often follow a monomodal distribution and the definition
of a partition, fuzzy or not, tends to be arbitrary.
To our knowledge, the implicative analysis has not yet been applied to
microarray data. However, the discovery of association rules has been recently
proposed in this particular application field.
In [8], association rules are extracted from gene expression databases rel-
ative to the yeast genome. A preprocessing retains the genes that are under-
expressed or over-expressed according to their expression values. This work is
based on the A priori algorithm [1] and the usual rule parameters, support and
 Microarray analysis 207

confidence. In [24], the authors present a set of operators for the exploration
of comprehensive rule sets. The expression values are discretized according
to predetermined thresholds. The rules are filtered with classical support and
confidence parameters. One drawback of these two methods is the dependency
of the obtained rules to arbitrary thresholds. A similar study [5] incorpo-
rates annotation information, combined with over- or under- expression. [20]
presents HAMB, a machine learning tool that induces classification rules from
gene expression data. FARMER [7] also discovers association rules from mi-
croarray datasets. Instead of finding individual association rules, FARMER
finds interesting rule groups, i.e. a set of rules that are generated from the same
set of individuals. FARMER uses a supervised discretization procedure, based
on entropy minimization. A case study on human SAGE data [3] explores
large-scale gene expression data using the Min-Ex algorithm, which efficiently
provides a condensed representation of the frequent itemsets. The data have
been transformed into a boolean matrix by a discretization phase, the logical
true value corresponding to gene over-expression. The authors analysed the
effect of three different discretization procedures. Our work is closer to the
one proposed in [16]. The authors define the concept of emerging patterns,
where itemsets are boolean comparison operators over gene expressions. They
use an entropy minimization criterion that strongly differs from our approach,
since it takes into account all the samples, while we prefer to extract higher
quality rules, even if they concern only a subset of observations.

3 Implication over rank intervals

3.1 Definitions

For the sake of generality, we consider a set of m individuals for which n mea-
surements have been performed. In our study, individuals are genes (actually
gene products to be more precise) and the measurements correspond to a set
O of n different experimental conditions. A set of experiments generally refers
to a biological study involving different tissue samples. We will call observa-
tion an experiment relative to a particular biological condition and implying
the whole set of individuals (genes). Let M (k, l) be the measurement value
associated to an indidividual k and an observation l. Note that this value may
refer to any ordinal data type. Our analysis relies on this matrix, although
the same study could be performed on the transposed matrix. We call profile
of the individual k the vector p(k) = (M (k, l), l ∈ [1, n]). The profile in our
application is usually called the expression profile and concerns the whole set
of measurements relative to gene k. We define the operator rank that takes
a profile p(k) and returns its observation indexes, ranked in increasing order.
For example, let us consider the profile p(k) = (4.1, 12.3, 1.2, 3.7). We have
rank(p(k)) = (3, 4, 1, 2), which means that the lowest value (i.e. 1.2) has been
measured under condition 3, the value 3.7 under 4, and so on. The study
208 Gerard Ramstein

of the observation ranks may reveal hidden relationships. In market basket

analysis, it can provide interesting relations between amounts of transactions.
We could for example discover association rules such as the fact that if a cus-
tomer buys a lot of pizza, its shopping cart will contain few fresh foods. In
microarray data, similar studies will concern associations between expression
levels. A rule A → B will for instance denote an association between an over-
expresssion observed on a gene A and an over-expression observed on a gene
B. Note that these types of interactions concern only a subset of observations
(e.g. a customer type in market basket data, experimental conditions in mi-
croarray data). We then need to precise a rank interval that limits the range
of the associations. Let I be the set of all subintervals of [1, n]:
I = {[p, q], (p, q) ∈ [1, n]2 , p ≤ q} (1)
By misuse of language we will simply call interval rank the observations
belonging to a given interval rank. Thus, the interval rank rk (i) will refer to
the set of observations relative to the interval i of I. This set is defined as
follows: rk (i) = {oj ∈ rank(k), j ∈ i}. In the previous numerical example, the
interval rank rk ([1, 2]) contains the observations 3 and 4, and corresponds to
the two lowest values of the profile. Table 1 gives an example of two profiles.
The visual inspection of the values does not permit to detect any relations
between the two individuals A and B. Table 2 gives a better insight of a
possible hidden association. After ranking, the observations 1, 4 and 9 are
close together in both observation sets. The previous definition helps us to
precise this similarity as follows: the rank intervals rA ([4, 6]) = {9, 1, 4} and
rB ([6, 9]) = {1, 3, 9, 4} are joint sets. This similarity can be verified by using
the statistical implicative analysis. This approach gives a quantitative answer
to the following question: is this set conjunction really surprising or could
comparable results be achieved by chance? The latter hypothesis corresponds
to the case where no hidden structure exists between the two profiles. Conse-
quently, the rank operator will not give any further information, which means
that the sets rA ([4, 6]) and rB ([6, 9]) could be considered as the result of a
random selection process. Implicative analysis offers a convenient framework
to address this problem. Following [15], we suppose two sets α and β having
the same respective cardinals than rA (i) and rB (j). The intensity of implica-
tion ϕ(α, β) represents the quality of the association of rA (i) → rB (j). In the
previous example, the intensity of implication equals to 0.86, which represents
an association of little significance.

profile 1 2 3 4 5 6 7 8 9 10
p(A) 6 4 10 7 8 13 3 12 5 2
p(B) 15 12 16 19 10 14 8 7 17 21
Table 1. Measurement values issued from individuals A and B.
 Microarray analysis 209

rank o1 o2 o3 o4 o5 o6 o7 o8 o9 o10
r(A) 10 7 2 9 1 4 5 3 8 6
r(B) 8 7 5 2 6 1 3 9 4 10
Table 2. Reordering of the observations using the rank operator. The values repre-
sent the observations of the previous table, namely the column indexes.

It is possible that other intervals of I exist for which a better conjunction

would be observed. It is therefore preferable to define the quality of a rule
A → B as the best association that can be found among all the possible
intervals. We then adapt the concept of intensity of implication as follows:

ϕI (A, B) = max(ϕ(rA (i), rB (j)), (i, j) ∈ I 2 ) (2)

Note that this measure is very robust with respect to the data: it is insen-
sitive to monotonic transformations, an interesting property for microarray
data, often prone to various preprocessings. The value of ϕI (A, B) indicates
the quality of the association. Besides, the relative intervals imax and jmax
for which the maximum defined in eq. 2 has been found provides useful in-
formation. The rule A → B can thus be expressed in a more precise and
operational form.
Let us define tA = min(M [A, o], o ∈ rA (imax)), TA = max(M [A, o], o ∈
rA (imax)), tB = min(M [B, o], o ∈ rB (jmax)) and TB = max(M [B, o], o ∈
rB (jmax)). Let o be an observation. The association rule can be expressed as
follows:
if tA ≤ M [A, o] ≤ TA , then tB ≤ M [B, o] ≤ TB (3)
Suppose for example that table 1 concerns two genes A and B issued from the
expression matrix M defined at the beginning of this section. The rule A → B
can then be written as follows: if for an observation o we have 5 ≤ M [A, o] ≤ 7,
then 15 ≤ M [B, o] ≤ 19.

3.2 Numerical and computational considerations

As definition 2 concerns an optimal value, the intensity of implication of coreg-

ulated genes often lies in [0.9, 1]. For practical reasons, it is easier to express
this quality measure as follows:

λI (A, B) = − log10 (1 − ϕI (A, B)) (4)

This measure presents the advantage to possess an infinite positive range and
to also increase with the quality of the rule. The parameter λI (A, B) is eas-
ily interpretable: instead of having for instance ϕI (A, B) = 0.9999, we will
consider λI (A, B) = 4, which means that the risk of observing a comparable
210 Gerard Ramstein

situation by chance is equal to 10−4 . In this paper we will equally use expres-
sion 2 or 4 in our numerical examples.

Definition 2 requires to explore all the subintervals of I. For one profile

k, we need to translate a window of size wk = q − p + 1 over the interval
[1, n − wk + 1]. For a simple rule A → B, the time complexity is O(n4 )
(without considering the rank operator, in O(nlog(n))). This search can be
limited by considering the following property: high quality rules necessarily
concern joint sets having similar sizes. Indeed, wA cannot be much larger
than wB , since if wA ≥ wB the number of counterexamples is always greater
than or equal to wA − wB . Reciprocally, if wB is much larger than wA , the
intensity of implication of the rule is very low. Therefore, it is more judicious to
consider a common window size w and to set wA = w and wB = w + , where
w ∈ [wmin , wmax ] and  ∈ [min , max ]. Then, wmin , wmax , min and max are
input parameters of the extraction algorithm. A more important algorithmic
restriction concerns the interval set I, that will be replaced by the following
set:
I 0 = {[p, q], (p, q) ∈ [1, n]2 , p ≤ q, p = 1 ∨ q = n} (5)
The interval set I 0 has a particular importance for microarray data, as bi-
ologists generally search for expression levels corresponding to either under-
expression or over-expression.
This is due to the fact that experimentation is mostly based on differential
situations. This means that the measurement defines the gene expression in
a tissue sample relatively to a reference sample. For example the study may
concern an ill tissue versus a sound one, or experiments on patients having
absorbed drug or not. The two extremities of the ranking are then the most
interesting observations related to “abnormal” gene activity. For this reason,
we propose to replace the computation of λI (A, B) by λI 0 (A, B). The rule
discovery algorithm using this new interval set is less time consuming (O(n2 )
complexity instead of O(n4 )) and provides a simpler interpretation of rules.
In the case of over-expressions for instance, a rule may be transposed using
eq. 3 in the following condensed form: if M [A, o] ≥ tA then M [B, o] ≥ tB .

3.3 Comparison of association rules and correlation techniques

The most widely used microarray analysis concerns the study of gene coreg-
ulation. A gene A is said to be coregulated with a gene B if the expression
profile of both genes is similar. This similarity is generally measured by differ-
ent metrics, such as the Euclidean distance or the Pearson’s correlation. Note
that the intensity of implication is oriented and can favour an association
from A to B rather than from B to A. The usual metrics do not precise any
orientation and has another drawback: the similarity measure is issued from a
global estimation that takes into account all the observations of O while our
implicative analysis searches for partial similarity between interval ranks that
 Microarray analysis 211

are subsets of O. A largely used technique in microarray analysis consists in

hierarchical clustering based on absolute Pearson correlation. As a clustering
method, this measure may not be the most pertinent one, as demonstrated
by the example shown in fig. 1. This figure represents two profiles issued from
two genes belonging to the cerevisiae specie, commonly known as bakers’yeast.
The yeast genome is essential for research in molecular biology since it cor-
responds to a relatively simple organism permitting numerous experiments.
As a considerable body of knowledge has been collected regarding the yeast
genes, it is a particularly interesting case study. Our example is issued from
the microarray dataset selected in [11]. It comprises a set of genes whose func-
tions are known and that do not contain missing values. The gene expression
has been measured under 89 different experimental conditions such as heat
shock or nitrogen starvation.
Figure 1 represents the implication of gene CHA1 over gene SAM1. The
filled triangles indicate that the former is under-expressed in response to a
signal of amino acid starvation, and, to a smaller degree, of nitrogen depletion.
CHA1 is involved in the catabolism of threonine, an essential amino acid
found in peptide linkage in proteins. Filled circles show that SAM1, a gene
interfering in the metabolism of the methionine, is over-expressed for the same
experiment set.
2
expression values

0
−2
−4
−6

0 20 40 60 80

observations

Fig. 1. Implication of CHA1 over SAM1. The absciss axis represents the 89 exper-
imental conditions. The ordinate axis represents the expression measures. Triangles
belong to the profile of gene CHA1 (YCL064C) and circles belong to the profile of
SAM1(YLR180W). The filled points denote the observations belonging to the rank
intervals that maximize the intensity of implication. The two rank intervals are iden-
tical and the corresponding rule accepts one exception, indicated by a double arrow.
This arrow points out to the fact that there exists an observation (shown by an
unfilled triangle) that is less under-expressed and that does not belong to the rank
interval of SAM1 while it is included in that of CHA1.

This particular subset of conditions covers only 9% of the set O. Cor-

relation measures are incapable to detect such partial associations. Table 3
summarises the values of different statistic measures, including the intensity
212 Gerard Ramstein

of implication for this pair of genes. These results show that the implicative
measure is finer than correlation techniques. Indeed the low values obtained
from the latter will not permit the gene association to emerge, notably with
respect to the large amount of data. In contrast with these results, the im-
plicative value clearly reveals the quality of the rule, stating that the risk to
encounter such an association by chance is less than one over a thousand.

method value
Intensity of implication 0.9992
Pearson correlation -0.16
Kendall correlation -0.0089
Table 3. Comparison of the implication and correlation measures.

4 Application to tumour discrimination

The microarray technology being complex and expensive, the vast majority
of the expression data correspond to supervised analysis for which the obser-
vations are relative to precise phenotypes. We will consider a very important
application, namely the classification of tumour samples. This challenging
problem refers to the assignment of particular tumour samples to already-
defined classes, based on gene expression monitoring by DNA microarrays.
These classes define different tumour subtypes. Recognizing these subtypes is
crucial to determine the malignity of the cancer. Depending on the subtype,
the clinical course can indeed vary from indolence over decades to explosive
growth and patient’s death.
We suppose the a priori knowledge of a set C of classes, each observation
belonging to only one class of C. In our application study, the observation
is relative to a patient for whom a precise diagnosis has been made. Let
L(ok ) = cj ∈ C be a class labelling function that associates to an observation
ok its corresponding class cj . We will address two related problems. The first
one concerns the prediction of the class corresponding to an unkown obser-
vation from a learning dataset. The second one is the selection of the most
discriminating genes, also called informative genes or marker genes, i.e. genes
whose expression profile is different from one class to the other. As it has
been explained in section 3.2, these differences generally correspond to either
under-expression or over-expression. The selection of informative genes is a
challenging problem, because of the great number of genes that are analysed
(from several thousands to the whole genome on a single chip). One observes
that most of the genes on a microarray do not participate to the discrimina-
tion of tumour types and that the majority of them present a low amplitude of
 Microarray analysis 213

variation. The definition of marker genes is crucial for clinical investigations,

since it permits to predict the outcome for a patient with a relatively low-cost
procedure. We first describe a gene selection method based on implicative
analysis and then its application to tumour classification.

4.1 Gene selection

To drastically reduce the number of genes, biologists generally use a filtering

technique which is mostly based on statistic tests such as the t-test, ANOVA-
F, Cochran, Kruskall-Wallis, Brown-Forsythe and Welsh [6]. The implicative
method can be applied to extract informative genes, by the discovery of clas-
sification rules of the form:
rg (i) → c (6)
This condensed notation expresses the fact that the observations o of rg (i)
verify L(o) = c. In this rule, conclusion c gives the class label associated to
gene g. The premise corresponds to the rank interval relative to i ∈ I 0 . We
consider the interval set I 0 defined in eq.5 to focus our analysis on either
under-expression or over-expression. In other words, the classification rule
states that the observed expression level on gene g and defined by the interval
i concerns observations belonging to class c.
As in eq. 2, the quality of a classification rule is given by the optimal
intensity of implication over interval ranks:

ϕI 0 (g → c) = max(ϕ(rA (i), Oc ), i ∈ I 0 , Oc = {o ∈ O | L(o) = c}) (7)

The only difference with eq. 2 is that rB (j) is now replaced by a unique obser-
vation set, the set Oc of observations of class c. As we search for classification
rules, we only consider predefined classes. Note however that a more complex
analysis could be performed by accepting any subset of O, this issue being
similar to the selection of genes in an unsupervised study. When the class par-
tition is unequal, the use of the intensity of implication presents an important
advantage. Contrary to the confidence, our measure takes into account the
fact that a class c is over-represented. Note that the intensity of implication
is always null in the extreme case where Oc = O.
The selection of informative genes among a gene set G is performed by the
following algorithm:
Selection algorithm
inputs:
M, G : expression matrix and its gene set
C, L : the classes and the class labelling function
K : the number of informative genes per class
outputs:
igs : the informative gene set
begin
214 Gerard Ramstein

igs ← ∅
for each class c ∈ C do
genelist ← ∅
for each gene g ∈ G do
– compute ϕ = ϕI 0 (g → c)
– genelist ← genelist ∪ {(g, ϕ)}
end;
– sort all pairs (g, ϕ) ∈ genelist in decreasing order of ϕ
– add into igs the K first genes of the sorted gene list
end;
end.
Note that the selection process permits to discover informative genes ca-
pable of discriminating more than one class among the others. A typical sit-
uation will concern genes presenting an over-expression for a given class and
an under-expression for another one. A special case is relative to datasets
that are partitioned into two classes. One expects that a gene that discrimi-
nates one class will automatically be informative for the other one. Actually,
this assertion will not necessarily be true, according to the position of coun-
terexamples at each extremity of the ranking. Two different triplets (g, c1 , ϕ1 )
and (g, c2 , ϕ2 ) will indeed be associated to a same gene g. In the case where
ϕ1 6= ϕ2 , the same gene may be retained in one class and rejected in the other.

Application to leukemia subtype discrimination

The study [12] presents a well-known experiment of cancer classification based
solely on gene expression monitoring. This paper has demonstrated that mi-
croarrays can provide a tool for cancer classification. This work explores the
capacity of gene expression analysis to dicriminate between two subtypes of
tumour, acute myeloid leukemia (AML) and acute lymphoblastic leukemia
(ALL). Differentiating ALL from AML is critical for successful treatment: it
has be proven that distincting therapies improves cure rates and diminishes
toxicities. We used the initial dataset of the authors, which contains measure-
ments corresponding to ALL and AML samples from bone marrow. These 38
samples concern 27 ALL and 11 AML. The dataset initially comprises 6,817
human genes. We proceeded to a filtering phase for the normalization of the
data and we eliminated genes whose official symbol name could not be re-
covered. After this preprocessing, we obtained a set of 3571 genes. Figure 2
shows the distribution of the intensity of implication λI 0 relative to the whole
set of classification rules that can be extracted. The histogram reveals a non-
negligible proportion of genes that are potentially discriminative. Approxi-
matively 10% of the gene set present an intensity of implication ϕ greater
than 0.9997 (λ ≥ 3.52).This means that around 300 genes have differential
expression with respect to ALL and AML subtypes.
This property is also implicitly mentioned in [12]. However, the authors
select 50 genes to represent the most informative genes for the discrimination
 Microarray analysis 215

350
300
250
200
150
100
50
0

0 2 4 6 8 10

Fig. 2. Histogram of the intensity of implication. The absciss axis represents λI 0

and the ordinate axis corresponds to the number of genes.

between ALL and AML. As we have two classes, we did the same by setting
the input parameter K of our selection algorithm to 25. We obtained a set of
classification rules that comprises 14 genes described by the authors as dis-
criminative and included in their list of 50 genes. Table 4 shows that the genes
selected by Golub et al. are comparable whith our set in terms of intensity of
implication. The same minimum has been found in both sets and the mean is
almost identical. However, one observes for our selected set that the dispersion
is lower, which seems to indicate that the quality of our classification rules is
higher. To verify this assumption and to analyse the discriminative power of

gene set min median max mean variance

−10 −6 −5 −6
Golub & al. 8.3 10 3.6 10 2.8 10 7.6 10 6.9 10−11
Our gene set 8.3 10−10 3.4 10−7 3.2 10−6 1.2 10−6 1.6 10−12
Table 4. Comparison of gene sets. The statistics indicated in this table refer to the
intensity of implication.

these two gene sets, we proceeded to a comparison of their prediction capacity.

We applied the K-Nearest-Neighbours algorithm for holdout validation tests
(K = 3). For each gene set G that we tested, we used the expression matrix M
reduced to the genes belonging to G. We thus obtained two distinct matrices,
one corresponding to the gene set of the authors and the other to our own
gene set. Table 5 shows that our selection provides better holdout validation
results.
216 Gerard Ramstein

% test Golub & al. our method p-value

50% 3.11 0.79 1.2e-7
25% 1.67 0.11 2.7e-4
10% 1.00 0.00 4.5e-02
2.6% 0.00 0.00 1.0
Table 5. Holdout validation in the Golub learning set. The first column gives the
percentage of the dataset that has been used as test set. The second and third
columns represent the mean error rate expressed in percentage over 100 random
sets. The p-value is obtained from a Student’s t test. The results indicate that the
error rate differences are statistically significant for small learning sets.

4.2 Tumour classification

Cancer remissions highly depend on specific therapies that distinguish the

treatments according to distinct tumour types. Cancer classification has his-
torically relied on specific biological insights. DNA microarray technology
permits to discriminate between tumour subtypes that present the same mor-
phological appearance. Tumour classification by gene expression monitoring
is then a crucial and challenging task.
The previous section has shown that the intensity of implication can be
used to determine the most informative genes. We now propose to examine
the quality of the classification rules for the prediction of tumour subtypes.
Our learning dataset comprises a set G of genes and a set O of observations,
providing an expression matrix M . All observations of O are classified by a
class labelling function L.
Let Γ (M ) be the set of classification rules that are extracted from this
learning dataset, using the selection algorithm described in the previous sec-
tion. The cardinal of Γ (M ) is K. | C |, where K is the input parameter of
the selection algorithm and | C | the number of classes. Let s be an unknown
tissue sample (i.e. a new observation) for which an expression profile p(s) has
been measured. This vector contains the expression values corresponding to
the gene set G of the learning set. We call pg (s) the expression value associ-
ated to gene g ∈ G. The classification procedure is defined by the following
algorithm:

Classification algorithm
inputs:
M : expression matrix
Γ (M ) : set of discriminative rules
p(s) : the expression profile of an unknown sample s
outputs:
cs : the predicted class of s
begin
parameter:
 Microarray analysis 217

count : vector of size | C |, initially set to 0

for each class c ∈ C do
for each rule (rg (i) → c) ∈ Γ (M ) do
– Let tg and Tg be resp. the minimal and
the maximal expression value relative to the gene g
and to the observation set rg (i)
if tg ≤ pg (s) ≤ Tg then
count[c] ← count[c] + 1
end;
end;
cs < −argmaxc∈C (count)
end.
Different microarray datasets are publicly available for tumour classifica-
tion studies. We illustrate our method with the following data:
Brain tumour dataset This dataset contains gene expression profiles from 5
different tumours of the central nervous system. 42 tumour tissue samples
are partitioned according to their tumour subtypes: 10 medulloblastomas,
10 malignant gliomas, 10 atypical teratoid/rhabdoid tumours (AT/RTs), 8
primitive neuro-ectodermal tumours (PNETs) and 4 human cerebella. The
raw data are available at the web site of the Whitehead Institute Center
for Genomic Research (http://www-genome.wi.mit.edu/cancer). After
prepocessing, 5,597 genes remained.
Colon cancer dataset This dataset contains expression levels of colon tissues.
The study concerns 40 tumoural and 22 normal tissues. The expression of
6,500 genes has been measured using the Affymetrix technology. The data
are available at the web site of the Colorectal Cancer Microarray Research
(http://microarray.princeton.edu/oncology).
Leukemia dataset This dataset comprises the expression of 72 tumours rela-
tive to acute lymphoblastic leukemia (ALL, 47 cases) or acute myeloid
leukemia (AML, 25 cases). The gene expressions were obtained from
Affymetrix oligonucleotide microarrays. The data are available at the web
site of the Whitehead Institute Center for Genomic Research
(http://www-genome.wi.mit.edu/cancer).
Table 6 summarizes the main properties of our test set.

Dataset Publication # samples # classes # genes Response

Brain Pomeroy [22] 42 5 5, 597 tumour subtypes
Colon Alon [2] 62 2 2, 000 tumoural/normal tissues
Leukemiea Golub [12] 72 2 3, 571 tumour subtypes
Table 6. Publicly available datasets.
218 Gerard Ramstein

We compare our method with two major contributions in tumor classifica-

tions that emphasize the critical importance of feature selection. [9] is based
on a supervised clustering of genes and a plurality voting with classification
trees. [26] uses Self-Organizing Maps and fuzzy c-means clustering. We also
tested the following general-purpose machine learning algorithms:
k-Nearest-Neighbours. This classification algorithm extracts the k nearest
neighbours of an unknown sample s, according to a distance function
d(x, y). We have used the absolute Pearson coefficient, its associated dis-
tance being expressed as follows:
Pn
(xi − µ(x))((yi − µ(y))
d(x, y) = 1− | i=1 | (8)
(n − 1)σ(x)σ(y)

where µ and σ are respectively the mean and the standard deviation
of the expression profile. The classification algorithm assigns s to the
most numerous class within the neighbour set. When many features are
bound to be little relevant, feature-weighted distances are preferable to
the Pearson distance. However, the standard k-NN method is easy to
implement and, compared to more sophisticated techniques, it provides
relatively good classification results for microarray data [27].
Random Forest. This classifier [4] consists of many decision trees that deal
with a random choice of samples (with replacement). At each selection
node, only a random choice of conditions is used. The forest selects the
classification having the most votes over all the trees in the forest. Random
forest is especially well-suited for microarray data, since it achieves good
predictive performance even when the number of variables is much larger
than the number of samples, as it has been demonstrated in [10].
Support Vector Machines. A support vector machine [25] is a machine lear-
ning algorithm that finds an optimal separating hyperplane between mem-
bers and non-members of a given class in an abstract space. As random
forests, this classifier shows excellent performance in high-dimensional
variable space and then is well-adapted to classification of microarray
samples [19].
Table 7 presents the results we obtained on a leave-one-out validation
test with these classifiers and our three datasets. Our method, although it is
based on a simple counting of the rules that s verifies, achieves comparable
performances with the most sophisticated techniques.

5 Analysis of gene association networks

In the previous section, we presented the use of implicative analysis for tumour
classification and the selection of informative genes. We will now illustrate the
application of association rules to the study of gene associations. The selection
 Microarray analysis 219

method Brain Colon Leukemia

our method 14.3 12.9 2.8
Gene clustering 11.9 16.1 2.8
Fuzzy c-means 14.3 11.4 4.1
random forest 19.0 14.5 2.8
support vector machine 11.9 12.9 2.8
K Nearest Neighbours 23.8 22.6 1.4
Table 7. Comparison of different classifiers. We give the percentage of error rates
relative to a leave-one-out validation procedure. The first line concerns the results
obtained from our classification algorithm (the classification rules are of course ex-
tracted from learning sets that do not contain the unknown sample s).

algorithm will enable us to focus our analysis on the most discriminative

genes. We will first propose an original visual representation of the gene space,
based on implicative analysis. We will then present a partial view of the gene
association network, corresponding to the most informative genes.

5.1 Gene representation based on implicative analysis

We apply our study on tumour samples relative to brain cancer. The dataset is
the same as the one presented in the previous section (42 samples, 5 classes).
It is the most complex dataset, because of its number of tumour subtypes
and its classification error rates. As in [4] the data are preprocessed. This
preprocessing comprises thresholding, filtering, a logarithmic transformation
and standardisation of each experiment to zero mean and unit variance. Fil-
tering includes a selection of the first thousand of genes by decreasing order
of variance.
The gene selection process extracts the K = 10 most discriminative genes
for each of the five tumour types. Let Γ (M ) be our new set containing these
50 genes. We compute for each pair (gi , gj ) of genes in Γ (M ) the intensity
of implication associated to gi → gj . We obtain a gene association network
that can be vizualised using any arbitrary layout algorithm. We prefer to
position the genes with respect to the quality of their associations with their
neighbours. Therefore we define a similarity function sim(gi , gj ) as follows:
sim(gi , gj ) = max(ϕ(gi , gj ), ϕ(gj , gi )). To visualise our genes, we express the
distance between two genes gi and gj as follows:

distance(gi , gj ) = ms − sim(gi , gj )

where ms is the maximal value of the elements of the matrix sim. Multi-
dimensional scaling (MDS [18]) provides a visual representation of the prox-
imities among a set of objects. This method permits to associate to a gene g
a point in a plane.
220 Gerard Ramstein

Figure 3a shows the mapping of the genes. It reveals a good clustering

of the genes according to their class label. The clusters are well separated
and correspond to the predefined classes, without any counterexamples. To
compare this result with standard techniques, we replaced our distance func-
tion by the absolute Pearson distance defined in eq. 8. We applied it to the
same set of genes and performed the MDS algorithm on the new distance
matrix. Figure 3b indicates that the Pearson correlation is not capable of
separating the genes into distinct clusters. This difference demonstrates that
the implicative method can better discriminate the gene similarity observed
on a particular subset of observations, contrary to correlation methods that
consider the observation set as a whole.
The application of the implicative method has proven to be efficient in the
particular context of the unsupervised analysis. This problem refers to class
discovery and concerns experiments for which the cancer subtypes are un-
known. In this case, the distance function that we propose can be directly
used to identify clusters. Therefore, it suffices to apply an efficient cluster-
ing algorithm such as PAM (Partionning Around Medioids [17]) that extracts
clusters from a dissimilarity matrix.

(a) (b)

Fig. 3. Representation of the gene similarity. (a) represents the MDS based on im-
plication analysis: the different tumour types associated to our gene selection are well
separated, contrary to (b), in which the dissimilarity measure is the absolute value
of the Pearson correlation. We obtained a similar mapping by using the Euclidian
distance instead of the Pearson coefficient (figure not shown).

5.2 Discovery of gene association using the intensity of implication

The previous model gives an original insight of the similarities between genes.
An interesting feature of implicative analysis has been neglected so far: its
 Microarray analysis 221

capability to define the orientation of these similarities. The implication is

indeed an oriented relation and reveals that an association A → B is more
or less pertinent than its counterpart B → A. It is necessary to apprehend
the meaning of implications over genes. An asymmetric relationship between
two genes traduces the fact that the knowledge of the expression value rel-
ative to one gene better determines the order of magnitude (i.e the interval
range) of the expression value on the other than the contrary. This property
can somehow be related to gene regulation: the activity of one gene causes
the change of expression rate of another. One must however be careful not
to interpret implication rules as the direct evidence of a gene regulation. The
microarray experiments cannot encompass all the biological mechanisms that
take place in the cell. Nevertheless, gene association networks may help bi-
ologists to appreciate the polarity of gene coexpressions. Figure 4 presents
the 20 genes belonging to the set Γ (M ) (K being set to 10) issued from the
Leukemia dataset presented in section 4.2.

MARCKSL1
NMB

LYN SNRPN
ELA2

TCF3
APLP2

CD63

CCT3
MGST1
CD33
TOP2B

CST3 ACADM

FAH
ZYX
RBBP4

NCOA6

ADM PPBP

Fig. 4. Gene association network

In a dataset comprising two classes, there exist only two types of gene
profiles that discriminate these classes. Following the analysis done by the
authors [12], we then consider the two following patterns:
π1 : Genes that present an under-expression for ALL samples and an
over-expression for AML samples.
222 Gerard Ramstein

π2 : Genes that present an over-expression for ALL samples and an

under-expression for AML samples.
It seems interesting to discriminate these two kinds of situations. Unfor-
tunately, definition 2 authorizes associations over gene profiles that are an-
ticorrelated, as in the example shown in figure 1. This means that we can
extract strong associations having either an under-expression in premise and
an over-expression in conclusion, or the opposite. In both cases, these associa-
tions concern genes belonging to distinct types of profiles. In order to simplify
the interpretation of the association network, we then introduce the following
restriction: we only retain associations rA ([pi , qi ]) → rB ([pj , qj ]) that respect
the condition: (pi = pj = 1) ∨ (qi = qj = n). Note that this simple constraint
does not assume the existence of classes.
One can observe in fig. 4 that two clusters emerge from the association
network. The left cluster corresponds to the differential expression π1 previ-
ously described, the right cluster to π2 .
The nodes of the network are the informative genes labelled by their official
symbol name. The edges of the network represent associations whose inten-
sity of implication is greater than 0.999 (λ(A, B) ≥ 3). One remarks that the
right cluster is less connected than the left one. This feature indicates that
the expression profile π1 is more observed than π2 . Although it is only a con-
jecture, this phenomenon could explain why, in section 4.1, we achieve better
performances in holdout validation with our selected genes, compared to [12]:
contrary to the authors, we did not impose the predefined expression patterns
π1 and π2 .
The association network presents several genes that are mentioned in
biomedical literature. For example, CD33 has been proven to be useful in
distinguishing lymphoid from myeloid lineage cells (see [12] for more details).
Genes that are the sources of many arrows (e.g. FAH and ZYX) have a pro-
file that is very close to their respective expression pattern. This means that
the measurements are well separated into two expression levels, each level
approximatively corresponding to one class of observation.
Table 8 gives some examples of rules that have been extracted and repre-
sented in fig. 4. The level column indicates whether the rule is relative to an
under-expression (interval ([1, q], 1 ≤ q ≤ n, noted `) or to an over-expression
([p, n], 1 ≤ p ≤ n, noted a). The subtype defines the class c of observations
that is mostly concerned by the rule. The pattern column indicates the type
of profile deduced from the two previous pieces of information. The support
is the percentage of observations of O that respect the rule. The class column
gives the percentage of observations of class c that are concerned by the rule.
The homogeneity h is the percentage of observations belonging to the major-
ity class c that respect the rule. This measure is the percentage expression of
 Microarray analysis 223

the sensitivity (also called the true positive rate). Finally, the rule quality λ
is the intensity of implication expressed in its logarithmic form (eq. 4).
The rule F AH → ZY X in line 5 is the most pertinent rule, since it con-
cerns all the individuals of class ALL (100 % of class support) and only them
(100% of class homogeneity): it is a perfect indicator of the leukemia subtype.
Rule ACADM → RBBP 4 (line 1) presents a high intensity of implication,
although it does not concern all the observations of class ALL. Indeed, 7.4%
of observations of class ALL do not respect the rule (note however that this
rule conversely presents a perfect class homogeneity, i.e. all the observations
belong to ALL class). The high value of λ is due to its support that is greater
than that of rule 5 (i.e. rule 1 concerns more individuals than rule 5). Rule 2
has the same premise than the previous rule. Its class homogeneity is reduced.
Almost one third of the individuals do not belong to the majority class of the
rule. This rule reflects the fact that coregulation is not always associated to
the leukemia subtype. It is indeed possible to encounter a remarkable statis-
tical association between genes that is not necessarily linked to the studied
phenotype. That is why one must consider all the parameters expressed in
table 8 before interpreting a rule.

line rule level subtype pattern support class h λ

1 ACADM → RBBP 4 ` ALL π1 65.8 92.6 100 4.2
2 ACADM → T CF 3 a AML π1 42.1 100 68.7 3.7
3 CD33 → CD63 ` AML π2 34.2 100 84.6 4.0
4 F AH → P P BP a ALL π2 68.4 96.3 100 3.6
5 F AH → ZY X ` AML π2 28.9 100 100 3.8
6 M ARCKSL1 → SN RP N a AML π1 28.9 90.9 90.9 3.2
7 M ARCKSL1 → T CF 3 a AML π1 23.7 81.8 100 3.0
8 T OP 2B → T CF 3 ` ALL π1 28.9 40.7 100 3.3
9 ZY X → LY N a ALL π2 57.9 81.5 100 3.3
10 ZY X → P P BP ` AML π2 26.3 90.9 100 3.4
Table 8. Association rules between genes.

6 Conclusion
Microarray data analysis is generally based on the measure of the similar-
ity between gene expression profiles, such as the absolute Pearson correlation.
The drawback of this measure is that it assumes a global relationship between
genes, while an implication may only concern a particular group of conditions.
The intensity of rank implication is more appropriate for the discovery of par-
tial dependencies. Association rule may therefore help to infer gene regulatory
pathways. Our method is very robust to noise and, unlike correlation tech-
niques, it provides the direction of the relationship.
224 Gerard Ramstein

We also propose a gene selection method based on implicative analysis. Our

model identifies informative genes that have proven very predictive, compared
to selection sets existing in the literature. Such a model is potentially useful
for medical diagnoses. Indeed, a reliable classification of tumours is essential
for successful treatment of cancer. Although simple, our rule-based classifier
is an efficient algorithm, which achieves performances comparable to high per-
forming machine learning techniques.
Due to their simplicity and ease of interpretation, association rules are very
promising for the analysis of gene expression data. They determine a net-
work of genes that are differentially or coordinately expressed under specific
conditions. The implicative analysis can be used to discover such association
networks.

References
1. R. Agrawal and R. Srikant. Fast algorithms for mining association rules. In Pro-
ceedings of the 20th Very Large Data Bases Conference, pages 487–499. Morgan
Kanfmann, 1994.
2. U Alon, N Barkai, D A Notterman, K Gish, S Ybarra, D Mack, and A J Levine.
Broad patterns of gene expression revealed by clustering analysis of tumor and
normal colon tissues probed by oligonucleotide arrays. Proc Natl Acad Sci U S A,
96(12):6745–6750, Jun 1999.
3. C. Becquet, S. Blachon, B. Jeudy, J. F. Boulicaut, and O. Gandrillon. Strong-
association-rule mining for large-scale gene-expression data analysis: a case
study on human sage data. Genome Biol, 3(12), 2002.
4. L. Breiman. Random forests. Machine Learning, 45:5–32, 2001.
5. Pedro Carmona-Saez, Monica Chagoyen, Andrés Rodríguez, Oswaldo Trelles,
José María Carazo, and Alberto D. Pascual-Montano. Integrated analysis of
gene expression by association rules discovery. BMC Bioinformatics, 7:54, 2006.
6. D. Chen, Z. Liu, X. Ma, and D. Hua. Selecting genes by test statistics. Journal
of Biomedicine and Biotechnology, 2:132–138, 2005.
7. G. Cong, A. Tung, X. Xu, F. Pan, and J. Yang. Farmer: Finding interesting
rule groups in microarray datasets, 2004.
8. C. Creighton and S. Hanash. Mining gene expression databases for association
rules. Bioinformatics, 19(1):79–86, January 2003.
9. M. Dettling and P. Buhlmann. Supervised clustering of genes. Genome. Biol.
Res., 3(12):research0069.1–0069.15, 2002.
10. R. Diaz-Uriarte and S. Alvarez de Andres. Gene selection and classification of
microarray data using random forest. BMC Bioinformatics, 7, 2006.
11. A. Gasch and M. Eisen. Exploring the conditional coregulation of yeast gene
expression through fuzzy k-means clustering, 2002.
12. T. R. Golub, D. K. Slonim, P. Tamayo, C. Huard, M. Gaasenbeek, J. P. Mesirov,
H. Coller, M. L. Loh, J. R. Downing, M. A. Caligiuri, C. D. Bloomfield, and E. S.
Lander. Molecular classification of cancer: class discovery and class prediction
by gene expression monitoring. Science, 286:531–537, 1999.
13. R. Gras, R. Couturier, F. Guillet, and F. Spagnolo. Extraction de règles en
incertain par la méthode statistique implicative. In 12èmes Rencontres de la
Société Francophone de Classification, pages 148–151, Montreal, 2005.
 Microarray analysis 225

14. R. Gras, E. Diday, P. Kuntz, and R. Couturier. Variables sur intervalles et

variables-intervalles en analyse statistique implicative. In Société Francoph-
one de Classification (SFC’01), pages 166–173, Univ. Antilles-Guyane, Pointe-
à-Pître, 2001.
15. R. Gras. L’implication Statistique. La Pensée Sauvage, Grenoble, 1996.
16. Xiaoming Jin, Xinqiang Zuo, Kwok-Yan Lam, Jianmin Wang, and Jia-Guang
Sun. Efficient discovery of emerging frequent patterns in arbitrary windows on
data streams. In ICDE, page 113, 2006.
17. L. Kaufman and P.J. Rousseeuw. Finding Groups in Data: an Introduction to
Cluster Analysis. Wiley-Interscience, New York, 1990.
18. J. B. Kruskal and M. Wish. Multidimensional Scaling. Sage Piblications, Beverly
Hills, CA, 1978.
19. Yoonkyung Lee and Cheol-Koo Lee. Classification of multiple cancer types by
multicategory support vector machines using gene expression data. Bioinfor-
matics, 19(9):1132–1139, 2003.
20. Gary Livingston, Xiao Li, Guangyi Li, Liwu Hao, and Jiangping Zhou. Analyz-
ing gene expression data using classification rules. In CSB’2003: Proceedings of
the 2003 Computational Systems Bioinformatics Conference, pages 1–8. ACM
Press, August 2003.
21. Gregory Piatetsky-Shapiro and Pablo Tamayo. Microarray data mining: facing
the challenges. SIGKDD Explorations, 5(2):1–5, 2003.
22. Scott L Pomeroy, Pablo Tamayo, Michelle Gaasenbeek, Lisa M Sturla, Michael
Angelo, Margaret E McLaughlin, John Y H Kim, Liliana C Goumnerova,
Peter M Black, Ching Lau, Jeffrey C Allen, David Zagzag, James M Olson,
Tom Curran, Cynthia Wetmore, Jaclyn A Biegel, Tomaso Poggio, Shayan
Mukherjee, Ryan Rifkin, Andrea Califano, Gustavo Stolovitzky, David N Louis,
Jill P Mesirov, Eric S Lander, and Todd R Golub. Prediction of central ner-
vous system embryonal tumour outcome based on gene expression. Nature,
415(6870):436–442, Jan 2002.
23. A Schulze and J Downward. Navigating gene expression using microarrays–a
technology review. Nat Cell Biol, 3(8):190–195, Aug 2001.
24. Alexander Tuzhilin and Gediminas Adomavicius. Handling very large numbers
of association rules in the analysis of microarray data. In KDD, pages 396–404,
2002.
25. Vladimir N. Vapnik. The nature of statistical learning theory. Springer-Verlag
New York, Inc., New York, NY, USA, 1995.
26. Junbai Wang, Trond Hellem Bø, Inge Jonassen, Ola Myklebost, and Eivind
Hovig. Tumor classification and marker gene prediction by feature selection
and fuzzy c-means clustering using microarray data. BMC Bioinformatics, 4:60,
2003.
27. C H Yeang, S Ramaswamy, P Tamayo, S Mukherjee, R M Rifkin, M Angelo,
M Reich, E Lander, J Mesirov, and T Golub. Molecular classification of multiple
tumor types. Bioinformatics, 17 Suppl 1:316–322, 2001.
On the use of Implication Intensity for matching
ontologies and textual taxonomies

Jérôme David, Fabrice Guillet, Henri Briand, and Régis Gras

Laboratoire d’Informatique de Nantes Atlantique

Equipe COnnaissances & Décision
Site Ecole Polytechnique de l’Université de Nantes
La Chantrerie — BP 50609 — 44306 Nantes cedex 3
{jerome.david,fabrice.guillet,henri.briand}@univ-nantes.fr,
[email protected]

Summary. At the intersection of data mining and knowledge management, we shall

hereafter present an extensional and asymmetric matching approach designed to find
semantic relations (equivalence and subsumption) between two textual taxonomies
or ontologies. This approach relies on the idea that an entity A will be more specific
than or equivalent to an entity B if the vocabulary (i.e. terms and data) used to
describe A and its instances tends to be included in that of B and its instances. In
order to evaluate such implicative tendencies, this approach makes use of association
rule model and Interestingness Measures (IMs) developed in this context. More
precisely, we focus on experimental evaluations of IMs for matching ontologies. A set
of IMs has been selected according to criteria related to measure properties and
semantics. We have performed two experiments on a benchmark composed of two
textual taxonomies and a set of reference matching relations between the concepts
of the two structures. The first test concerns a comparison of matching accuracy
with each of the selected measures. In the second experiment, we compare how each
IM evaluates reference relations by studying their value distributions. Results show
that the implication intensity delivers the best results.

Key words: Ontology alignment, ontology matching, association rule, interesting-

ness measure

1 Introduction
In our information society, large databases and data warehouses have become
widespread. This huge amount of information has led to the increasing de-
mand for mining techniques for discovering knowledge nuggets. To meet this
demand, the Knowledge Discovery in Databases (KDD) [16] community pro-
posed the association rule model [1].
Initially motivated by the analysis of market basket data, the task of asso-
ciation rule mining aims at finding relations between items in datasets [9].
J. David et al.: On the use of Implication Intensity for matching ontologies and textual
taxonomies, Studies in Computational Intelligence (SCI) 127, 227–245 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
228 J. David et al.

Association rules are propositions of the form “If antecedent then conse-
quent”, noted antecedent → consequent, representing implicative tendencies
between conjunctions of valued attributes or items. Association rules have the
advantage of being an easy and meaningful model for representing explicit
knowledge. Furthermore, this unsupervised learning technique does not need
particular information about knowledge to be discovered contrary to classical
supervised techniques (such as decision trees). These advantages have moti-
vated a great deal of research and the publication of association rule extraction
algorithms such as Apriori [1, 2]. Nevertheless, if only minimal support and
confidence values are used, these algorithms typically produce many rules and
it is hard to only select those which may interest the user. One way to face this
problem is to use Interestingness Measures (IMs). IMs aim at assessing the im-
plicative quality of association rules but also some useful characteristics such
as novelty, significance, unexpectedness, nontriviality, and actionability [9,19].
IMs allow to rank and reduce the amount of rules, and consequently to help
the user to choose the best ones according to his/her preferences.
The information society has also led to the development of the Web and
then a great increase in the available data and information. In this vast
Web, resources, often in textual form, tend to be organised into hierarchi-
cal structures. This hierarchical structuring of web contents ranges from large
web directories (e.g. Yahoo.com, OpenDirectory) to online shop catalogs (e.g.
Amazon.com, Alapage.com). Furthermore, with the arrival of the Semantic
Web, such a hierarchical organisation is also used through OWL ontologies
which aim at providing formal semantics of the web contents. Even if the use
of hierarchies helps to structure web information and knowledge, the Web re-
mains heterogeneous. Data exchanges and communications between software
programs or software agents using hierarchical-organised data is consequently
difficult. In order to address such interoperability problems, one must be able
to compare such data structures and find matches between them. Thus, many
matching methods have been proposed in the literature [27, 34, 35]. These
methods aim at finding semantic relations (i.e. equivalence, subsumption, etc)
between entities (i.e. directories, categories, concepts, properties) defined in
different hierarchical structures (filesystems, schemas, ontologies). Even if the
proposed approaches are issued from different communities, they mostly use
similarity measures and as a consequence, a majority of them are restricted
to finding equivalence relations only.
At the intersection of these two research fields, we proposed to use the as-
sociation rule paradigm for matching ontological structures [11]. Our original
approach, named AROMA (Association Rule Ontology Matching Approach),
heavily relies on the asymmetric nature of association rules, which allows it to
match not only equivalence relations but also subsumption relations between
entities. The consideration of subsumption between entities helps to charac-
terise more precisely the matching relations between hierarchical structures
regarding only similarity based approaches. Also, it allows enhancements the
output matches. Furthermore, unlike most approaches designed for matching
 On the use of II for matching ontologies and textual taxonomies 229

schema or ontologies, AROMA relies heavily on extensional data provided

with structures. This type of matcher, named an instance-level matcher [34],
is especially designed to work on structures with limited schema informa-
tion (i.e. only concept or element names and a partial order relation between
them). For example, AROMA can deal with textual hierarchies, such as Web
directories or semi-structured data.
The main objective of this chapter is to show and explain the relevance of
the Implication Intensity measure relative to other IMs in the context of the
problem of matching text hierarchies.
This chapter is organised as follows: in a first section, we present re-
lated work concerning ontology/schema matching and propose a classification.
Then, we focus on IMs proposed in the literature. First, we classify measures
according to three criteria. Then, we use this classification to justify the selec-
tion of the best IMs according to our matching context. In the second section,
we detail the two stages of AROMA methodology and, we describe a criterion
for reducing rule redundancy and enhancing the accuracy of matching results.
The last section reports the results of two experiments made on a well-known
benchmark provided with two catalogs and a set of reference matching rela-
tions. The first experiment evaluates the selected IMs according to a classical
information retrieval accuracy measure: the F-measure. In the second exper-
iment, we compare the value distributions of IMs obtained on relevant and
non-relevant relation sets.

2 Related work
2.1 Textual taxonomy matching
For the last six years, ontology and schema matching have been widely studied
and many approaches have been proposed in the literature. These methods
come from different communities such as artificial intelligence [7,18,20], data-
bases [12, 29, 32], graph matching [23, 30], information retrieval [28], machine
learning [13, 31, 36], natural language processing and statistics [24]. Although
they are heterogeneous and consequently difficult to compare, preliminary
afforts have been result in surveys of matching techniques [27, 34, 35].
One survey [34] focuses on database schema matching techniques and pro-
poses a classification which discriminates the extensional or element-based ap-
proaches from the intensional or only-schema-based approaches. While many
efforts are concentrated around the intensional matchers, few extensional ap-
proaches have been proposed in the literature.
A survey of intensional matchers can be found in [35]. The authors pro-
pose two classifications of this type of matchers. The first one permits us to
distinguish methods according to their granularity and their interpretation of
the input information (it distinguishes element-level and structure-level tech-
niques and then the syntactic, external and semantic methods). The second
classification distinguishes three classes:
230 J. David et al.

1. the terminological approaches (T) based on string similarity measures

(TS) or based on measures which use terminological resources (TL) like
WordNet;
2. the structural approaches (S) which compare two concepts from their in-
ternal structure (SI) (shared attributes or properties) or from their ex-
ternal structure (SE), that is to say their respective position within their
taxonomy;
3. the semantic approaches (SEM), which use formal semantic model.
To sum up, most approaches use terminological matchers with string-
similarities (Anchor-PROMPT [32], Coma [12], CMS [26], Cupid [29],
S-MATCH [20]) or/and external oracles such as Wordnet ( [20], H-match [8]).
They can also use structural matchers (Similarity Flooding [30], [8], GMO [23],
Omap [12,18,29,32,36]). Some methods such as OLA [15], combine the various
terminological and structural criteria. Only a few approaches use formal
semantic models [18, 20].

Extensional matchers

In this section, we focus on extensional matchers since the methods designed

for matching textual taxonomies (i.e. web directories, catalogs, etc.) heavily
rely on textual content. Beacause, such data structures typically have poor
schema information, extensional matchers are relevant. We briefly present four
approaches which are explicitly designed to work on such data structures.
Then, we propose a synthetic classification of these approaches based on their
differences.
GLUE [13]. This conceptual hierarchy matching tool uses machine learn-
ing techniques. It combines two strategies: the estimation of joint probabilities
between concepts and the relation labeller. In the first strategy, for the joint
probability estimation, the two hierarchies must share the same set of textual
documents. As it is often not the case, the authors propose to classify the
documents associated to the concepts of the first hierarchy to the concepts of
the second one. This classification process uses several machine learning clas-
sifiers: a naive Bayesian classifier on the textual content of the documents,
and a naive Bayesian classifier on the names of concepts concatenated with
the names of ascendant concepts (concept path). These two classifiers are
firstly trained on the documents associated with each hierarchy. The different
classifier predictions are then combined by using the meta-learner. The joint
probability is evaluated using Jaccard similarity.
oPLMap [31]. This tool is based on a logical and probabilistic model:
DATALOG. It aims at finding the matching set which maximises the align-
ment probability. This method considers rules of the form Si → Tj and eval-
uates the confidence of the rules by combining several classifiers. It relies on
terminological classifiers for concept names (concept name identity, Jaccard
measure on words composing concept names, Jaccard measure by considering
 On the use of II for matching ontologies and textual taxonomies 231

the concept path name) and machine learning classifiers for textual docu-
ments (kNN classifier and Naive Bayes text classifier on the documents). The
oPLMap approach also uses some constraints for taking into account the struc-
ture of the taxonomies. In the output, this method provides a set of n-to-n
mapping elements valued by a probability measure.
Hical or SBI [25]. This method uses a statistical test, named κ-
statistic [10], on shared documents for determining matching. κ-statistic tests
the null hypothesis: “κ = 0”. A relation between concepts holds if the null
hypothesis can be dissmissed with a significance level of 5%. Hical proposes a
top-down approach in order to reduce the computing time.
CAIMAN matching service [28]. This matching method is enclosed
in the CAIMAN system for facilitating the exchange of relevant documents
between geographically dispersed people within their communities of inter-
est. In the CAIMAN matching service, each document is represented by a
document vector composed of words and word frequencies weighted by the
TF/IDF measure. Then, the characteristic vectors of concepts are computed
from the document vectors by using the Rocchio classifier. Finally, similarities
between concepts are deduced by evaluating the cosine value between char-
acteristic vectors of concepts. The output matching is a one-to-one matching:
for each source concept, the method retains only the target concept for which
the measure is maximised.

Classification of textual taxonomy matchers

Kind of rela- Measure Comparison Pre-

tions level processing
step
GLUE equivalence Jaccard document level classification of
documents
oPLMap equivalence confidence document level classification of
(conditional documents
probability)
Hichal equivalence k -statistic document level none
CAIMAN equivalence cosines term level characteristic
vectors compu-
tation
AROMA subsumption, Implication In- term level selection of
equivalence tensity sets of relevant
terms
Table 1. Comparison of hierarchy matchers
232 J. David et al.

Table 1 compares of four matchers and AROMA method discussed in

Section 3. We analyse them according to four criteria:
• The kind of relation that the method considers. It can be equivalence or
subsumption relations.
• The measure used by the method for evaluating a potential matching be-
tween concepts. The measures used in the five approaches are Jaccard
similarity, the conditional probability (called confidence in the association
rule community), κ-statistic, cosine and the Implication Intensity.
• The comparison level shows how concepts are represented and then ex-
plains on which basis the similarity or measure is calculated. We denote
two modalities: (1) the document level for methods comparing shared doc-
uments between entities; and (2) the term level for methods which repre-
sent entities by a set of terms.
• The kind of pre-processing step refers to the kind of preprocessing applied
to the extension in order to be able to calculate similarities. This charac-
teristic is partially determinated by the type of comparison level. In the
case of document level approaches, documents can be classified from each
structure into the other one so that they share the same set of documents.
Term level approaches use linguistic processing in order to extract terms
and then select and/or weight these terms according to their relevance to
the studied entities.
We can see that only AROMA considers the subsumption relation. The
others are restricted to the equivalence. Concerning the measures used, three
methods use similarities (which are symmetric) while only oPLMap and
AROMA use asymmetric measures. Nevertheless, oPLMap does not seem to
use its asymmetric measure to find subsumption relations between concepts.
We can also note that Hichal and AROMA use measures based on a statis-
tical model. For the comparison level, three methods use the documents for
representing the extension of concepts from which they calculate the measure
values. Only CAIMAN and AROMA work at the terminological level by rep-
resenting concepts by sets of terms (characteristic vectors or relevant term
sets). Finally, in order to work on a common document base, the GLUE and
oPLMap methods use a combination of classifiers for the documents.

2.2 Interestingness Measures

In the framework of association rule discovery, and in order to select the most
interesting rules, many Interestingness Measures (IMs) have been proposed
and studied (see [3, 19, 22, 37] for a survey). In this context, some researchers
are interested in principles and properties defining a good IM [3,17,22,33,37],
while others work on the comparison of IMs from a data-analysis point of
view.
According to our objective of hierarchy matching, we selected some IMs
that may be relevant for our work. In the context of AROMA, unlike
 On the use of II for matching ontologies and textual taxonomies 233

association rule discovery, an IM is not used for ranking rules in a post-

processing step, but during the rule extraction process. In order to be able
to choose a threshold value more easily, we retained only IMs respecting the
principle of minimal and maximal value [22]. IMs respecting this behaviour
are more intelligible to a user.
In this section, we firstly introduce notations used for representing associ-
ation rules and their characteristics. Then, we classify selected IMs according
to three main criteria proposed in [4,6]. The resulting taxonomy of IMs shows
their main properties and permits understanding of their behaviours and se-
mantics.

Definition of association rule and notations

In this section we use the following notation: A finite set T of n individuals

is described by a set I of p items. Each transaction t can be considered as
an itemset, so that t ⊆ I. A = {t ∈ T ; a ⊆ t} is the extension of itemset a
and B = T − {t0 ∈ T ; b ⊆ t0 } is the extension of b. An association rule [1]
is an implication of the form a → b, where a and b are disjoint itemsets. In
practice, it is quite common to observe some transactions which contain a and
not b without in spite of a general trend to have b when a is present. Then, we
introduce the quantities na = card(A), nb = card(B) and na∧b = card(A∩B).

Taxonomy of IMs

The taxonomy classifies IMs according to three criteria. The first one concerns
subject of IMs (deviation from independence or equilibrium) and the second
one, the nature of IMs (descriptive or statistical). The last one, the scope
of the IMs (quasi-implication, quasi-conjunction, quasi-equivalence), explains
the semantics of the measure according to logical operators.

• Subject. A family of IMs evaluate the deviation from the independence

situation where the number of counter-examples is equal to those expected
in a random case (nab = na .nb /n). These measures have a fixed value at
the independence. The other family of IMs evaluate the deviation from
the equilibrium. The equilibrium situation is reached when the number of
counter-examples and examples are equal in number (nab = nab ).
• Nature. The nature of an IM can be descriptive or statistical. The descrip-
tive measures are not influenced by a proportional expansion of the cardi-
nalities taken into account. A descriptive IM m satisfies m(na , nb , nab , n) =
m(α.na , α.nb , α.nab , α.n) with α. > 0. Conversely, the statistical measures
vary with the expansion of cardinalities. According to the authors, this
type of IM allows the validity of rules to be statistically assessed. Some
of these measures are also particularly effective at detecting rules that
have novel consequents, i.e. consequents that have not been seen in pre-
vious rules. For example, Implication Intensity decreases with an increase
234 J. David et al.

of nb , and thus gives preference to statistically valid rules having novel

consequents.
• Scope. Finally, this last way of distinguishing IMs relies on the idea that
IMs may evaluate a proximity between the rule and a logical configuration
such as an implication, a conjunction, or an equivalence [4]. To qualify the
scope of an IM, we will use the terms quasi-implication, quasi-conjunction
and quasi-equivalence indexes because rules are not strict logical proposi-
tions since they may have counter-examples. Furthermore, some IMs only
evaluate the tendency to verify the consequent when the antecedent is true
(i.e. they only consider the examples of a rule). Such measures (e.g. the
Confidence or IPEE [5] measures) are not classified in the Table 2.
For each modality of the scope characteristic of an IM, Table 2 shows
their symbol, counter-examples, equivalent linkage and property that the IM
must respect. By comparing this table and the semantic relations defined
by [20], we can see that based on their scope, measures could be more or
less adapted for mining certain types of semantic relations. Subsumption rela-
tions must be evaluated by quasi-implication measures, overlapping relations
by quasi-conjunction measures and equivalence relations by quasi-equivalence
measures. In schema or ontology matching, methods often rely on quasi-
conjunction measures for evaluating equivalence relations of the approaches
considered in Section 2.1, only Hichal uses an index (κ or kappa) that is a
quasi-equivalence measure. GLUE uses a quasi-conjunction index (Jaccard)
and oPLMap uses a rule index (confidence or conditional probability).

quasi-implication quasi-conjunction quasi-equivalence

Symbol ⇒ ↔ ⇔
Counter- a∧b a∧b a∧b
examples
a∧b a∧b
a∧b
equivalence a⊃b≡b⊃a a↔b≡b↔a a↔b≡a↔b
IM’s property I(a → b) = I(b → a) I(a → b) = I(b → a) I(a → b) = I(a → b)
Table 2. Scope of IMs

Table 3 shows, for each selected IM, its scope (rule (→), quasi-implication
(⇒), quasi-conjunction (↔), quasi-equivalence (⇔)), its nature (statistical (S)
or descriptive (D)), its subject (deviation from independence (I) or equilibrium
(E)), its fixed value in the independence or equilibrium situation (depending
on its subject) and its formula.
 On the use of II for matching ontologies and textual taxonomies 235

Measure Scope Nature Subject Fixed Formula

value
na .n
II ⇒ S I 0.5 P (n
ab
< P oisson(
n
b ))
n.n
Loevinger ⇒ D I 0 1− ab
na .n
b
IPEE → S E 0.5 P (n
ab
< Binomial(na , 1/2))

Confidence → D E 0.5 nab /nb

na .nb
Likelihood ↔ S I 0.5 P (nab ≥ P oisson(
n
))

Linkage
Analysis
Table 3. Selected IMs and their properties

3 AROMA methodology
AROMA (Figure 1) was designed to find matching between conceptual hierar-
chies populated from textual documents. This method permits the discovery
of a set of significant association rules holding between concepts obtained
from two hierarchical structures and evaluated by the Implication Intensity
measure. AROMA takes, as input, two conceptual hierarchies H1 and H2 ,
each defined as a tuple H = (C, ≤, D, σ), where C is the set of concepts, ≤
is the partial order organising concepts into a taxonomy, D is the set of tex-
tual documents, and σ is the relation associating a set of documents to each
concept (i.e. for a concept c ∈ C, σ(c) represents the documents associated
to c). Thanks to the first part of the method concerning the acquisition and
the selection of relevant terms for each concept, we are able to redefine each
hierarchy as a tuple H0 = (C, ≤, T, γ) where T is the set of relevant terms
selected. In order to consider the partial order, we assume that a term associ-
ated with a concept is also associated with itsSparent concepts, and thus, we
extend γ to the relation γ 0 as follows: γ 0 (c) = c0 ≤c γ(c0 ).

3.1 Association rules discovery between hierarchies

The second stage of AROMA consists of the discovery of implicative match-

ing relations between concepts by evaluating association rule between their
respective sets of relevant terms. The algorithm takes in two pre-processed
hierarchies H0 1 and H0 2 and considers only the terms shared by the two struc-
tures. The set of common terms for the two hierarchies H0 1 and H0 2 is noted
0
T1∩2 = T1 ∩ T2 . The relation γ1∩2 associates a subset of T1∩2 for each concept
c ∈ C1 ∪ C2 :
n γ 0 (c) ∩ T if c ∈ C
0 1 2 1
γ1∩2 (c) = (1)
γ20 (c) ∩ T1 if c ∈ C2
The extracted rules are of the form a → b and are valued by a ϕ(a → b)
value. A valid rule a → b (i.e. rule having a ϕ(a → b) value greater than or
equals to a chosen threshold) represents a quasi-implication (i.e. an implication
allowing some counter-examples) from the set of relevant terms of the concept
236 J. David et al.

Fig. 1. The AROMA approach

a into the set of relevant terms of the concept b. The existence of such a valid
rule means that the concept a (issued from H1 ) is probably more specific than
or equivalent to the concept b (issued from H2 ).

3.2 Selection of significant rules

The algorithm provides a top-down search of association rules. It uses two

criteria for selecting significant rules and reduce redundancy. Then, a rule
a → b (between the concepts a ∈ C1 and b ∈ C2 ) will be significant if it
respects the two following criteria:

ϕ(a → b) ≥ ϕr (2)

∀x ≥ a, ∀y ≤ b, ϕ(x → y) ≤ ϕ(a → b) (3)

The first criterion (Equation 2) guarantees the quality of the implication
tendency between the two concepts for a given threshold ϕr . The Implication
Intensity of the rule a → b is explained in Figure 2 and defined as follows:


ϕ(a → b) = 1 − Pr Na∧b ≤ na∧b (4)
On the use of II for matching ontologies and textual taxonomies 237

Fig. 2. Implication Intensity of a rule a → b

Fig. 3. Evaluation of rules

238 J. David et al.

where na∧b = card(γ1∩2 (a) − γ1∩2 (b)) is the number of relevant terms for
concept a that are not relevant for concept b. Na∧b is the random number of
relevant terms for concept a that are not relevant for concept b.
For example (Figure 3), the rule A2 → B4 has nA2∧B4 = 1 counter-
examples. Its Implication Intensity value is calculated using a Poisson law
(which is a possible model for the Implication Intensity [21]):
nA2∧B4
X λk
ϕ(A2 → B4) = e−λ × = 0, 97
k!
k=0

where λ = nA2 .nB4 /n = 6.(30 − 8)/30.

The second criterion (Equation 3) verifies the generativity of the rule and
thus allows the redundancy to be reduced in the extracted rules set. Indeeds,
a valid rule (i.e. a rule satisfying the first criterion) is significant if there does
not exist a more generative rule having an Implication Intensity value greater
than or equals to it. A rule x → y is more generative than a rule u → v if
u ≤ x and y ≤ v (with x → y 6= u → v). For example (Figure 3), the rules
A2 → B7, A2 → B8, A1 → B4, A1 → B7, and A1 → B8 are more generative
than the studied rule A2 → B4. The rule A2 → B4 will be significant and
thus selected if none of its generative rules have a ϕ value greater than or
equals to its ϕ value.

4 Experimental results
The experiments presented in this section concern only the second part of
the AROMA (i.e. the rule selection phase). After describing the data used for
the experiments, we first compare the performance of the measures in terms
of the F-Measure, which aggregates precision and recall. Next, we describe
an analysis of the distribution of the measure values on two sets of matching
relations: a set of hand-made reference matching relations, noted R+ and a
set of irrelevant relations R−.

4.1 Analysed data

The experiments used the “Course catalog” benchmark [14]. This benchmark
is composed of two catalogs of courses descriptions which are offered at the
Cornell and Washington universities. The courses descriptions are hierarchi-
cally organised into schools and colleges and then into departments and centers
within each college. These two hierarchies contain respectively 166 and 176
concepts to which are associated 4360 and 6957 textual course descriptions.
The benchmark data also include a set of 54 manually matched relations from
concepts of the Cornell catalog to the Washington catalog. Only equivalence
relations are included in the manually matched set.
 On the use of II for matching ontologies and textual taxonomies 239

4.2 Evaluation of IMs

Here, we describe the evaluation performed by the AROMA algorithm with

each selected IM. For each measure, we varied the rule selection threshold
ϕr from 0 to 1 with a step of 1 percent. For each threshold value, the rule
selection algorithms was executed twice, once to select implications from the
Cornell concepts to the Washington concepts and the second time to select
implications from the Washington concepts to the Cornell concepts. From
the two implicative matching sets, we retained only equivalence relations by
following this rule: if A → B and B → A, then A ↔ B.
In order to evaluate the relevance of results according to the reference
set R+, we use two standard metrics from information retrieval: the preci-
sion and the recall. These metrics are defined as follows: let F be the set
of matching pairs found using AROMA and R+ be the set of “reference”
matching pairs. The precision (precision = card(F ∩ R+)/card(F )) mea-
sures the ratio of the number of good matching pairs (i.e. matching pairs
that are both in our result set and in the reference matching set) over
the number of matching pairs found by AROMA. The recall (recall =
card(F ∩ R+)/card(R+)) measures the ratio of good matching pairs over
the number of reference matching pairs. Finally, these two measures are ag-
gregated into the F-measure (Dice’s similarity between the sets F and R):
F − measure = 2.precision.recall/precision + recall.
Figure 4 shows a high correlation between Confidence and Loevinger in
terms of efficiency. Their best F-measure scores (around 0.45) are obtained
around a threshold of 0.3. This threshold value is greater than those of inde-
pendence situation for Loevinger but it is less than those of the equilibrium
situation for Confidence. The IPEE index has tendency to have the same
trend as Confidence and Loevinger. It has lower maximum F-measure values
but it is more robust to the increase of the selection threshold value. Likeli-
hood Linkage Analysis (LLA) don’t have good F-measure scores. It obtains
nearly constant recall values just above 0.5 but it has very bad precision val-
ues. Possibly the symmetric nature of this index is not well adapted to our
algorithm. Finally, the best F-measure scores are obtained by the Implication
Intensity index. Its best value is a little greater than 0.5 for a rule selection
threshold fixed at 0.9. We can also notice a stagnation of F-measure before
the independence situation (rule selection threshold of 0.5 for II).
These results show that Implication Intensity is the most relevant measure
in this context. The two descriptive indexes, Confidence and Loevinger, tend
to have similar trends. The selected quasi-conjunction index, LLA, is not
relevant for AROMA.

4.3 Distributions of IMs

In this experiment, we studied and compared how the IM’s evaluated matching
relations independently of the AROMA rule selection algorithms and their
240 J. David et al.

Fig. 4. Evolution of F-measure

selection criteria. Then, we propose to draw the values distributions of IM

in two cases. The first case consisted of evaluating the relations from the
manually matched set R+ and in the second case, we have tried the measures
on the set of irrelevant relations R−. The set R− was built manually and
contains the same number of relations as the reference one R+. For these two
evaluations, we performed the selection of relevant terms with the Implication
Intensity measure for a threshold of 0.9.
In the first test, for each manually matched relation represented by a pair
(A, B), we evaluated the rules A → B and B → A. For each pair, we kept only
the best value, in the cases where the studied measure was not symmetric.
Regarding Figure 5, the Confidence measure yields a value under 0.5 for
the majority of the relations. According to the equilibrium situation, these
relations are not relevant because their antecedents is more concomitant with
the negation of their consequents than with their consequents. The results ob-
tained with IPEE confirm this situation since many rules have a quality value
near 0. Nevertheless, the results found with Loevinger show that matching
relations are good regarding the independence case. The majority of rules
have a Loevinger value greater than 0 (only 4 relations have smaller values).
 On the use of II for matching ontologies and textual taxonomies 241

Distribution of Implication Intensity Distribution of Likelihood Linkage Analysis Distribution of Confidence

25

35

10
30
20

8
25
15
Frequency

Frequency

Frequency
20

6
15
10

4
10
5

2
5
0

0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Measure Value Measure Value Measure Value

Distribution of Ipee Distribution of Loevinger

15

10
8
10
Frequency

Frequency

6
4
5

2
0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Measure Value Measure Value

Fig. 5. Measures distributions on manual matching relations R+

Thus, we can say that their number of observed counter-examples is negatively

correlated to the expected one under the independence hypothesis.
The two probabilistic measures of deviation from independence, the Impli-
cation Intensity (II) and the Likelihood Linkage Analysis measures, evaluate
the majority of matching relations with good values. The second measure
seems to work a little better on this benchmark. This result is not surprising
because this second measure is symmetric and designed for evaluating quasi-
conjunction relations. But, unlike Implication Intensity, such a measure is not
adequate in the case of mining quasi-implication relations between concepts.
Then, after the study of distribution of measures values on good matching
relations, we performed a second test on a set of non-relevant relations. In
this case, we consider the minimal value obtained for each pair of concepts.
In Figure 6, all Confidence and IPEE values are near 0. Loevinger yields less
than 0 for 30 rules, that is to say under the independence situation. II confirms
this tendency because only 2 rules have a value greater than or equal to 0.5.
Nevertheless, LLA evaluates a majority of rules with good values (i.e. greater
than 0.5 obtained at the independence situation). Regarding Figures 5 and 6,
only Loevinger and II clearly distinguish the two sets of rules. These two sets
of distributions show that IMs of deviation from independence work better
242 J. David et al.

Distribution of Implication Intensity Distribution of Likelihood Linkage Analysis Distribution of Confidence

10

35
30
30
8

25
6
Frequency

Frequency

Frequency

20
20

15
4

10
10
2

5
0

0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Measure Value Measure Value Measure Value

Distribution of Ipee Distribution of Loevinger

30
25
30

20
Frequency

Frequency
20

15
10
10

5
0

0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Measure Value Measure Value

Fig. 6. Measures distributions on irrelevant matching relations R−

on this type of rules. We can also notice that the quasi-conjunction measure,
LLA, does not distinguish good rules from bad ones.
From these experiments, we conclude that matching relations are better
evaluated by IMs of deviation from independence. In such cases, the number
of counter-examples needed to reach the equilibrium situation is less than the
number needed to reach the independence situation. We also found that the
statistic measure of quasi-implication, II, is well suited for distinguish good
rules from bad ones.

5 Conclusion
In this paper, we proposed an original use of the association rule model and
interestingness measures in the context of schema/ontology matching. More
precisely, we described the AROMA approach, which is an extensional matcher
for hierarchies indexing text documents. A novel feature of AROMA is that it
uses of the asymmetrical aspect of association rules in order to discover sub-
sumption matches between hierarchies or ontologies. Based on studies of IMs,
we selected several IMs according to three criteria (subject, nature and scope)
and we evaluated them on a matching benchmark. The two experiments show
 On the use of II for matching ontologies and textual taxonomies 243

that deviation from independence measures are the best adapted IM family
for such an application since the evaluated rules are good regarding the in-
dependence situation, but bad in terms of equilibrium deviation. From these
results, we can also argue that the two descriptive indexes used, Confidence
and Loevinger, tend to have the same behaviour. Due to its deviation from
independence subject, its statistical nature and its quasi-implication scope,
the Implication Intensity obtains the best scores on this benchmark. In this
paper, we analysed measure behaviour on the process of rule extraction be-
tween concepts. We did not study the terminological step of AROMA, which
consists of extracting and selecting concept relevant terms. Such an evaluation
would be interesting since this terminological extraction process significantly
influences the accuracy of the results.

References
1. R. Agrawal, T. Imielinski, and A. Swami. Mining association rules between
sets of items in large databases. In Proceedings of the 1993 ACM SIGMOD
International Conference on Management of Data, pages 207–216. ACM Press,
1993.
2. R. Agrawal and R. Srikant. Fast algorithms for mining association rules. In
J.B. Bocca, M. Jarke, and C. Zaniolo, editors, Proceedings of the 20th Interna-
tional Conference Very Large Data Bases (VLDB’94), pages 487–499. Morgan
Kaufmann, 1994.
3. Jr. Bayardo, J. Roberto, and R. Agrawal. Mining the most interestingness
rules. In Proceedings of the 5th ACM SIGKDD International Conference On
Knowledge Discovery and Data Mining (KDD’99), pages 145–154, 1999.
4. J. Blanchard. A visualization system for interactive mining, assessment, and
exploration of association rules. PhD thesis, University of Nantes, 2005.
5. J. Blanchard, F. Guillet, H. Briand, and R. Gras. Assessing rule interestingness
with a probabilistic measure of deviation from equilibrium. In Proceedings of the
11th international symposium on Applied Stochastic Models and Data Analysis
(ASMDA-2005), pages 191–200. ENST, 2005.
6. J. Blanchard, F. Guillet, R. Gras, and H. Briand. Using information-theoretic
measures to assess association rule interestingness. In Proceedings of the fifth
IEEE international conference on data mining ICDM’05, pages 66–73. IEEE
Computer Society, 2005.
7. S. Castano, V. De Antonellis, and S. De Capitani Di Vimercati. Global viewing
of heterogeneous data sources. IEEE Transactions on Knowledge and Data
Engineering, 13(2):277–297, 2001.
8. S. Castano, A. Ferrara, and S. Montanelli. Matching ontologies in open net-
worked systems: Techniques and applications. Journal on Data Semantics,
3870(V):25–63, 2006.
9. A. Ceglar and J. F. Roddick. Association mining. ACM Computing Surveys,
38(2):5, 2006.
10. J. Cohen. A coefficient of agreement for nominal scales. Educational and Psy-
chological Measurement, 20(1):37–46, 1960.
244 J. David et al.

11. J. David, F. Guillet, R. Gras, and H. Briand. Conceptual hierarchies matching:

an approach based on discovery of implication rules between concepts. In Pro-
ceedings of the 17th European Conference on Artificial Intelligence (ECAI-2006),
pages pages 357–361, 2006.
12. H.H. Do and E. Rahm. Coma - a system for flexible combination of schema
matching approaches. In Proceedings of the 28th International Conference on
Very Large Data Bases (VLDB ’02), pages 610–621, 2002.
13. A. Doan, J. Madhavan, P. Domingos, and A. Halevy. Learning to map between
ontologies on the semantic web. In Proceedings of the 11th International WWW
Conference (WWW’02), pages 662–673. ACM Press, 2002.
14. A. Doan, J. Madhavan, P. Domingos, and A. Halevy. Ontology matching: a
machine learning approach. In S. Staab and R. Studer, editors, Handbook on
Ontologies in Information Systems, pages 397–416. Springer-Velag, 2004.
15. J. Euzenat and P. Valtchev. An integrative proximity measure for ontology
alignment. In Proceedings of the Semantic Integration Workshop, 2nd Interna-
tional Semantic Web Conference (ISWC-03), 2003.
16. U. M. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy, editors.
Advances in Knowledge Discovery and Data Mining. AAAI/MIT Press, 1996.
17. Alex Alves Freitas. On rule interestingness measures. Knowledge-Based Systems,
12(5-6):309–315, 1999.
18. F. Fürst and F. Trichet. Axiom-based ontology matching. In Proceedings of the
3rd international conference on Knowledge capture (K-CAP ’05), pages 195–196.
ACM Press, 2005.
19. Liqiang Geng and Howard J. Hamilton. Interestingness measures for data min-
ing: A survey. ACM Comput. Surv., 38(3):9, 2006.
20. F. Giunchiglia, P. Shvaiko, and M. Yatskevich. S-match: an algorithm and an
implementation of semantic matching. In European Semantic Web Symposium,
LNCS 3053, pages 61–75, 2004.
21. R. Gras et al. L’implication statistique, une nouvelle méthode exploratoire de
données. La pensée sauvage, 1996.
22. Robert J Hilderman and Howard J Hamilton. Knowledge Discovery and Mea-
sures of Interestingness. Kluwer Academic Publishers, 2001.
23. W. Hu, N. Jian, Y. Qu, and Y. Wang. Gmo: A graph matching for ontologies.
In Proceedings of the K-CAP 2005Workshop on Integrating Ontologies, pages
41–48, 2005.
24. R. Ichise, M. Hamasaki, and H. Takeda. Discovering relationships among cata-
logs. In E. Suzuki and S. Arikawa, editors, Proceedings of the 7th International
Conference on Discovery Science (DS’04), volume 3245 of LNCS, pages 371–379.
Springer, 2004.
25. R. Ichise, H. Takeda, and S. Honiden. Integrating multiple internet directories
by instance-based learning. In G. Gottlob and T. Walsh, editors, Proceedings of
the eighteenth International Joint Conference on Artificial Intelligence (IJCAI-
03), pages 22–30. Morgan Kaufmann, 2003.
26. Y. Kalfoglou and B. Hu. Cms: Crosi mapping system - results of the 2005
ontology alignment contest. In Proceedings of the K-CAP 2005 Workshop on
Integrating Ontologies, pages 77–85, 2005.
27. Y. Kalfoglou and M. Schorlemmer. Ontology mapping: the state of the art.
Knowledge Engineering Review, 18(1):1–31, 2003.
 On the use of II for matching ontologies and textual taxonomies 245

28. M. S. Lacher and G. Groh. Facilitating the exchange of explicit knowledge

through ontology mappings. In Proceedings of the 14th International Florida
Artificial Intelligence Research Society Conference (FLAIRS’01), pages 305–309.
AAAI Press, 2001.
29. J. Madhavan, P. A. Bernstein, and E. Rahm. Generic schema matching with
cupid. In Proceedings of the 27th International Conference on Very Large Data
Bases (VLDB’01), pages 49–58, 2001.
30. S. Melnik, H. Garcia-Molina, and E. Rahm. Similarity flooding: A versatile
graph matching algorithm and its application to schema matching. In Proceed-
ings of the 18th International Conference on Data Engineering(ICDE’02), pages
117–128. IEEE Computer Society, 2002.
31. H. Nottelmann and U. Straccia. A probabilistic, logic-based framework for
automated web directory alignment. In Zongmin Ma, editor, Soft Computing
in Ontologies and the Semantic Web, Studies in Fuzziness and Soft Computing,
pages 47–77. Springer Verlag, 2006.
32. N. Noy and M. Musen. Anchor-prompt: Using non-local context for seman-
tic matching. In Proceedings of the Workshop on Ontologies and Information
Sharing at the 17th International Joint Conference on Artificial Intelligence
(IJCAI’01), pages 63–70, 2001.
33. G. Piatetsky-Shapiro. Discovery, analysis and presentation of strong rules. In
G. Piatetsky-Shapiro and W. Frawley, editors, Knowledge Discovery in Data-
bases, pages 229–248. AAAI Press/MIT Press, 1991.
34. E. Rahm and P. A. Bernstein. A survey of approaches to automatic schema
matching. The VLDB Journal, 10(4):334–350, 2001.
35. P. Shvaiko and J. Euzenat. A survey of schema-based matching approaches.
Journal on Data Semantics, 4(LNCS 3730):146–171, 2005.
36. U. Straccia and R. Troncy. omap: Combining classifiers for aligning automat-
ically owl ontologies. In Proceedings of the 6th International Conference on
Web Information Systems Engineering (WISE-05), number 3806 in LNCS, pages
133–147. Springer Verlag, 2005.
37. Pang-Ning Tan, Vipin Kumar, and Jaideep Srivastava. Selecting the right ob-
jective measure for association analysis. Information Systems, 29(4):293–313,
2004.
Modelling by Statistic in Research
of Mathematics Education

Elsa Malisani and Aldo Scimone and Filippo Spagnolo

G.R.I.M. (Gruppo di Ricerca sull’Insegnamento delle Matematiche), Department

of Mathematics, University of Palermo, via Archirafi 34, 90123 Palermo (Italy)
http://www.math.unipa.it/~grim
[email protected], [email protected], [email protected]

Summary. The aim of this paper is to study the quantitative tools of the research
in didactics. We want to investigate the theoretical-experimental relationships be-
tween factorial and implicative analysis. This chapter consists of three parts. The
first one deals with the didactic research and some fundamental tools: the a priori
analysis of a didactic situation, the collection of experimental data and the statistic
analysis of data. The purpose of the second and the third section is to introduce the
experimental comparison between the factorial and the implicative analysis in two
researches in mathematics education.

Key words: research in didactics, theory of didactic situation, statistics, implica-

tive analysis, factorial analysis

Introduction

Modelling, by means of a statistical argumentation, supplies research in the

didactics of mathematics with a greater possibility of transferability of the
made experience.
It is evident, as it has been widely debated by La Casta-Brousseau [7],
Gras [14] and Spagnolo [11,32,34,36], that the statistical argumentation would
have not any valence without a theoretical remark from the view-point of
didactics and therefore of the epistemology of the mathematical contents.
Only a parallel study of all the possible argumentative paths of a research can
bring us to results which are considered reliable.

1 The research in didactics, some tools

The research in Didactics places itself as a goal-paradigm with respect to

other research paradigms in education science by using both the paradigm of

E. Malisani et al.: Modelling by Statistic in Research of Mathematics Education, Studies in

Computational Intelligence (SCI) 127, 247–276 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
248 E. Malisani et al.

the discipline subject of the analysis- and the paradigm of the experimental
sciences. The research in Didactics can be considered a sort of “Experimental
Epistemology”.
The fundamental tool for the research in didactics is the a priori analysis
of a didactic situation.
What does “a-priori analysis” of a didactic situation mean?
It means the analysis of the “Epistemological Representations”, “Historic-
epistemological Representations” and “Supposed Behaviours”, correct and not,
to solve a given didactic situation.

1. The epistemological representations are the representations of the possible

cognitive 1 paths regarding a particular concept. Such representations can
be prepared by a beginner subject or by a scientific community in a specific
historic period.
2. Historic-epistemological representations are the representations of the pos-
sible cognitive paths regarding the syntactic, semantic and pragmatic 2 re-
construction of a specific concept.
3. The supposed behaviours of students in facing the situation/problem are
all the possible strategies 3 of solution of it both correct and not. Among
the erroneous strategies, those which can become correct strategies will be
taken into consideration.

The a-priori analysis of the didactic situation allows:

1. the identification of the “space of the events” 4 regarding the particu-
lar didactic situation with respect to the professional knowledge of the
researcher- teacher in a specific historic period;
2. the identification, by means of the possible space of the events, of the
“good problem5 ” and therefore of a “fundamental didactic situation” for
the set of problems which the didactic situation refers to;

1
The cognitive paths permit the highlighting of the conceptual networks regarding
the didactic situation.
2
The semiotic perspective for the analysis of the disciplinary knowledge allows the
management of the contents with reference to the problems of “communication”
of the contents. This position is not particularly new with respect to the hu-
man sciences, but it represents a real innovation for the technical and scientific
disciplines.
3
In any case, a didactic situation poses a “problem” for the student to solve, either
as a traditional problem (i.e. in the scientific or mathematical framework) or as
a “strategy” to organise the best knowledge to adapt oneself to a situation.
4
For “space of the events” we mean the set of the possible strategies of solutions,
correct or not, and always supposed within a specific historic period by a specific
community of teachers.
5
The “good problem” is the one which, with respect to a given knowledge, allows
the best formulation in ergonomic terms.
 Modelling by Statistic in Research of Mathematics Education 249

3. the identification of the variables of the situation/problem and of the

didactic variables6 ;
4. the identification of the hypotheses of the research in Didactics of a more
general type with respect to those analysable by a first analysis of the
situation/problem.
So, the a-priori analysis represents the basic element of an educational
research and it takes into account both the epistemology of the discipline and
its history. Let us briefly summarise those parts of the research in didactics
which can be more significative, by pointing out other possible deeper studies
(Table 1) [6, 11].

Fig. 1. The diagram summarizes the relationship between research in didactics and
collection of experimental data.

1.1 The data

Each didactic research inevitably causes us to collect some data which can
be considered formed by a collection of elementary informations. Each piece
6
The “variables of a didactic situation” are all the possible variables which happen
into the situation. The “didactic variables” are those which permit a change of
the pupils’ behaviours. So, the didactic variables are a sub-set of the variables of
the didactic situation.
250 E. Malisani et al.

What is the • Research with its Paradigm;

Research in • Research with its own language;
Didactics? • Theoretical Research: Epistemological and
historic-epistemological analysis of the discipline with respect
to a specific Knowledge;
• Experimental Research:
1. A-priori analysis of the situation/problem;
2. Identification of the hypotheses of Research;
3. Falsification of the hypotheses7 ;
4. Analysis of the experimental data with respect to small
samples by means of appropriate statistical tools.
5. A-posteriori analysis of the experimental data.
Which is • Prevision of “didactic phenomena” through “Reliable models”
the use of • with respect to Theoretical-Experimental Research. For
the research “Reliable models” we mean those Models which allow the
in Didactics of possibility of making forecasts about didactic phenomena;
mathematics? • Communication of the results of the Research to the community
of Teachers by means of a strong argumentation such as the
a-priori analysis and the statistical tools.
What is •Problems regarding the “communication of a specific discipline”
Research in • through:
Didactics - Preparation of appropriate a-didactic situations;
concerned - Analysis of the errors and obstacles derived from the
with? communicative processes;
- Study of didactic and epistemological 8 obstacles such as:
• Tools for reflection on the construction of didactic curricula;
• Tools for a deeper and better understanding of the
communicative processes;
• Tools for the preparation of a-didactic situations.

Table 1.

of elementary information reports, in general, a behaviour of a pupil in a

given situation. A statistics, therefore, will be a set composed by: a student,
a situation and the behaviour of the student.
The student belongs to an observed sample E, assumed to be extracted
from a larger population, either by chance or following a system of control situ-
ations (for example: scholastic level, gender, previous personal knowledge. . . ).
The situation is chosen in a set S (of questions, exercises. . . ) generated and
structured by conditions and parameters of various natures (the knowledge in
play, material conditions, didactic conditions. . . ).
The behaviours (typical of knowledge or of aimed knowledge) are taken in
a set C of the student’s possible answers in the conditions which he is placed
into.
A class can be defined as a set E of students, a mathematics course as a
set S of exercises, the results of the students as a certain application of E onto
 Modelling by Statistic in Research of Mathematics Education 251

the set S × C where C is the set of the behaviours of successes or errors, a

note as an application of S × C onto R.
The knowledge of a certain behaviour can be represented by a certain
application of a set of questions onto a set of behaviours.

The use of statistics from teachers and researchers

The teacher must take rapid and many decisions and can correct them very
quickly if they prove to be inappropriate. He cannot wait for the result of the
statistical treatment of all his questions. The teacher must try to utilise these
statistical treatments which allow him to arrive quickly at certain conclusions.
The researcher must follow an opposite process:
1. Which hypotheses correspond to the questions that interest us?
2. What data should be collected?
3. Which treatments should be used?
4. What conclusions?

More than the rapidity and the immediate usefulness, it is the consistency,
the stability, the pertinence, and the sureness of the responses which interest
a researcher.
The research with appropriate statistical methods will allow:
1. the communication between teachers about the information they need and
which they collect on the results of the students; the value of the methods
used. . . ;
2. the use, also with discernment, of the results of the research in didactics;
3. the knowledge of the possibilities and the limits of the statistical methods
and so the legitimacy of the knowledge used in their profession;
4. the discussion about this legitimacy;
5. the formulation of some open conjectures to be put to the test of the
experimental contingency;
6. the imagination of the plausibility of these conjectures;
7. to know how to convert their experience into knowledge;
8. the participation in some research.

The teacher must take rapid and many decisions and can correct them very
quickly if they prove to be inappropriate. He cannot wait for the result of the
statistical treatment of all his questions. The teacher must try to utilise these
statistical treatments which allow him to arrive quickly at certain conclusions.
The researcher must follow an opposite process:
1. Which hypotheses correspond to the questions that interest us?
2. What data should be collected?
3. Which treatments should be used?
4. What conclusions?
252 E. Malisani et al.

5. More than the rapidity and the immediate usefulness, it is the consistency,
the stability, the pertinence, and the sureness of the responses which in-
terest a researcher.
6. The research with appropriate statistical methods will allow:
7. the communication between teachers about the information they need and
which they collect on the results of the students; the value of the methods
used. . . ;
8. the use, also with discernment, of the results of the research in didactics;
9. the knowledge of the possibilities and the limits of the statistical methods
and so the legitimacy of the knowledge used in their profession;
10. the discussion about this legitimacy;
11. the formulation of some open conjectures to be put to the test of the
experimental contingency;
12. the imagination of the plausibility of these conjectures;
13. to know how to convert their experience into knowledge;
14. the participation in some research.

Observations

An observation consists in attributing a value to a variable with reference to

an individual: the observed subject.
The statistics principally permits to treat the case when:
1. many observations are collected.
2. and when these observations totally:
a) regard different individuals with the same property;
b) refer to different properties for the same individual;
c) regard both cases i) and ii).
If “24” (value observed) is attributed to student X (the subject under
observation), as “result of the mathematical exam” (variable observed), the
set of the values or of the possible cases, in our case, is a whole number
between 0 and 30.
The variables can be numeric, interval, ordinal, nominal.
1. Numeric Variable: when the values are expressed by numbers (belonging
to the sets N, Z, Q, R) and the operations are significative for the variable.
2. Interval Variable: when only the interval among the values is significative
while the sum is not. For example, the points obtained in a sport can
constitute an interval variable.
3. Ordinal Variable: when the values express only an order among the ob-
servations. In an ordinal variable the sum of the two values is not a value.
4. Nominal Variable: When the values are characters (letters) or attributes.
This variable can have two values: its attribute and its negation. Even
if it is expressed by numbers, such as 0 and 1, a nominal variable is not
a numeric one: the sum between the two characters is not defined and
 Modelling by Statistic in Research of Mathematics Education 253

in general their order too. The only operations are the logical ones (set
theory).
It is always possible to transform a numeric variable into an interval, ordinal
or nominal variable (losing some information); an interval variable can be
transformed into an ordinal or nominal variable; an ordinal variable can be
transformed into a nominal variable. The reverse is not true.

1.2 The correspondence factor analysis and the implicative

analysis among variables in research in didactics of mathematics:
an experimental comparison

The research in didactics uses quantitative and qualitative tools. In this paper
we are trying to deal with the quantitative tools, spending more time, overall,
on theoretical-experimental relationships between factorial and implicative
analysis. Some significant experimental situations will be analysed in the parts
2 and 3.

Implicative analysis

The problem faced by R. Gras [12] arose from the attempt to answer the
following question: “Given some binary variables a and b, how can I be sure
that into a population, from each observation of a, there necessarily follows
the observation of b? ” Or in an even more succinct manner: “Is it true that
if a then b? ”
The answer is not generally possible and the researcher must be satisfied
with an “almost” true implication. By the implicative analysis by R. Gras, one
tries to measure the degree of validity of an implicative proposition between
binary and not binary variables. This statistical tool is used in the Didactics
of Mathematics9 .

Some observations on Factorial Analysis

The approaches to factorial analysis are of two types; the first one through the
study of the self-values of equations and the second one through a geometric
interpretation (vectors) and some contents of rational mechanics.
The approach presented here is the second one.
Let us consider the Cartesian product E × V (E constituted in general by
n students, n∈N; and V by m variables, m∈N). This is a typical situation
to find data in didactics. The problem is to represent geometrically the dis-
tribution of the two sets in a space of n × m dimensions. Factorial analysis
interprets the geometrical representations. This fact, in the sphere of Human
Sciences, has had many applications in the field of Psychology, but allowing an
9
All the information related to the mathematical theory are found in this volume.
254 E. Malisani et al.

analysis of small samples in the field of nonparametric Statistics significantly

contributes to the interpretation of didactic phenomena. See Bastin et al. [3]
for a geometrical approach to factorial analysis and Escofier and Pages [10]
for the interpretation of the graphic representation of the data.

1.3 Experimental comparison between Statistic Implicative

Analysis (SIA) and the Correspondence Factor Analysis (CFA)

The comparison between the two statistical tools has a prejudicial

epistemology:
1. Factorial analysis is in the field of descriptive statistics: averages, distances
by geometric methods. The measure used among variables is symmetrical.
This distance accentuates the role of rare observations. The geometrical
representations are given on different planes.
2. Implicative analysis is in the field of inferential statistics. The measure
used among variables is asymmetrical and highlights the not-commonplace
observations. The geometrical representations are carried out on a single
plane.
The types of variables treated are more varied in the ASI than in the CFA,
and one thinks they may be also fuzzy variables [37].
As regards the control of the data, the information:
1. in the ASI is complete and it is controlled by the indices of implication
and by the level of significance of the hierarchical levels:
2. in the CFA is controlled by the explicit inertia. There is a series of in-
dications to keep in mind to accept the level of the explicit inertia: for
example, regarding the number of observed variables or the number of
individuals.

Notwithstanding these two great epistemological differences, the two statisti-

cal tools are used for the research in the didactics of Mathematics as appro-
priate tools for the multi-varied analysis of small samples.
It is clear that the Correspondence Factor Analysis (CFA) can lead to the
study of variables (factors) which can be used to analyze experimentally the
conceptions, and the relationships among the conceptions referred to the iden-
tification of the factors. One has another thing, instead, when supplementary
variables are introduced. In the case of research in didactics the supplementary
variables are individuals made up in function of the hypotheses and of a-priori
analysis. On the transposed matrix the variables become the students with the
addition of the supplementary variables (theoretical students who correspond
to well defined characteristics by the nature of the problem of the research).
The consequent CFA, by the information with regards to the possible iden-
tified factors, gives information on how many individuals group themselves
onto one supplementary variable and this leads us to infer a reasoning of
 Modelling by Statistic in Research of Mathematics Education 255

the implicative type: If n individuals group themselves onto a supplementary

variable, then the supplementary variable identifies a significant conception.
The Statistical Implicative Analysis is concerned exactly with implications
and this activity is supplied without the introduction of supplementary vari-
ables. However, when the supplementary variables are introduced, implicative
analysis turns out to be much clearer and evident.
The comparison between the two methods can be significative only in the
case of the introduction of supplementary variables for what has been said up
to now.
Two experimental researches, on the same sample and with the same sup-
plementary variables, separately analysed, are compared.

Supplementary variables and Goldbach’s conjecture

In the doctoral thesis by Aldo Scimone, through the introduction of supple-

mentary variables regarding some reasoning schemes on the solution of Gold-
bach’s conjecture, the same data were compared by CHIC and by SPSS for
the factorial analysis10 .

The case of the passage from arithmetic to algebraic language

Elsa Malisani’s doctoral thesis compared the aspects of variable as unknown

and the functional relation in problem-solving, by considering the semiotic
contexts of algebra and analytical geometry. The goal was to investigate if
the notion of unknown interferes with the interpretation of the functional
aspect, and if the procedures in natural and/or arithmetic language prevail
as solving strategies for want of an adequate knowledge of algebraic language.
Also for this work we are trying to analyse the differences between factorial
and implicative analysis11 .

1.4 Conclusions

Both Correspondence Factor Analysis and Statistic Implicative Analysis do

analyses on small samples with the difference that CFA is a descriptive sta-
tistics while ASI is of inferential type. This first difference places ASI in a
different situation to infer from the sample of the population. However, as
already said in paragraph 2, the introduction of supplementary variables can
allow us an experimental comparison between the two statistical tools.
The measure used by CFA is symmetrical while the measure introduced
by ASI is asymmetrical. The asymmetry is due to the introduction of the
relationship of inclusion and so for its inferential nature (from the cause to
the effect).
10
See section 2.
11
See section 3.
256 E. Malisani et al.

The numerous data collected in experimental work regarding degree theses

and doctoral theses bring us to affirm that implicative analysis, when supple-
mentary variables are introduced, proves to be more incisive in the literature
of the contingencies. Instead, factorial analysis can give a useful argumentative
contribution as a support to implicative analysis when it must analyse only
the observed variables. This fact perfectly agrees the epistemological analysis
of the two statistical tools. So, ASI allows us a better approach to the sta-
tistical analysis of the evolution of the conceptions in the dynamics of the
classes.

2 The importance of supplementary variables in a case

of an educational research
2.1 The framework of the research

The theoretical framework of this educational research is the theory of didac-

tical situations in mathematics by Guy Brousseau [6].
It is known that this theory is based on the conception of the didactic sit-
uations, and in particular, this paper concerns an a-didactic situation, namely
that part of a didactic situation which teacher’s intention respect to pupils
is not clear into. An a-didactic situation is really the moment of the didactic
situation in which the teacher does not declare the task to be reached but he
gets the pupil to think about the proposed task which is chosen in order to
allow him to acquire a new knowledge which is to be looked for within the
same logic of the problem.
So, an a-didactic situation allows a pupil to appropriate and to manage
the staking dynamics, to get him to be a protagonist of the process, to get
him to perceive the responsibility of it as a knowledge and not as a guilt of
the sought result.

2.2 The historical context of the research

The research concerns some conceptions of pupils facing a conjecture, and in

particular the famous historical Goldbach’s conjecture. Goldbach’s conjecture
was chosen because it has a long historical background allowing an efficient
a-priori analysis, which is an important phase for the experimentation in order
to foresee the possible pupils’ answers and behaviours in front of the conjec-
ture. Moreover, it has a fascinating formulation allowing pupils to mix many
numerical examples, and to discuss fruitfully about its validity and some pos-
sible attempts for a demonstration of it. So, the historical context is important
because it suggests an interplay between the history of mathematics and the
mathematics education.
 Modelling by Statistic in Research of Mathematics Education 257

2.3 Using implicative analysis

This research was carried out by a quantitative analysis along with a qualita-
tive analysis. The statistical survey for the quantitative analysis was made by
two phases: in the first experiment, which was realized by a sample of pupils
attending the third and fourth year of study (16–17 years) of secondary school,
the method of individual and matched activity was used; the second experi-
ment was carried out in three levels: pupils from the first school (6–10 years),
pupils from primary school (11–15 years) and pupils from secondary school.
The quantitative analysis of the data drawn from pupils’ protocols was
made by the software of inferential statistics [14] CHIC 2000 (Classification
Hiérarchique Implicative et Cohésitive) and the factorial statistical survey
S.P.S.S. (Statistical Package for Social Sciences).
The research pointed out some important misconceptions by pupils and
some knots in the passage from an argumentative phase to a demonstrative
one of their activity which need to be deepened.

2.4 The experimentation

The research was realized on different levels by two experiment. The first ex-
periment, was realized with pupils attending the third and fourth year of study
(16–17 years) of secondary school, the method of individual and matched ac-
tivity was used. Pupils working individually were expected, within two hours,
to answer the following question:
a) Using the enclosed table of primes, the following even numbers can be
written as a sum of two primes (in an alone or in a manner more)? 248; 356;
1278; 3896.
b) If you have answered the previous question, are you able to prove that
it occurs for every even number?
The pupils working in couples were expected, within an hour, to answer
this question (in a written form and only if they agreed):
Is it always true that every even natural number greater than 2 is a sum of
two prime numbers? Let argue about the demonstrative processes motivating
them.
In both cases the procedure was acoustically recorded and the transcript
of those records with comments was made.
The second experiment was carried out in three levels: pupils from the
primary school (6–10 years), pupils from middle school (11–15 years) and
pupils from higher secondary school. The experiment was carried out on the
lowest level in two phases: In the first phase the pupils could answer this
question:
How can you obtain the first 30 even numbers by putting together prime
numbers of the table you have just made?
In the second phase, the pupils created small groups and tried to answer
the following question:
258 E. Malisani et al.

Can you derive the even numbers obtained by summing always and only
two primes? If it is so, can you state this is always the case for an even
number?
The pupils from lower secondary school solved the following problem
within 100 minutes:
Is the following statement always true? “Can an even number be resolved
into a sum of prime numbers?” Argue your claims.
The procedure had four phases:
a) a discussion about the task in couples (10 min.)
b) an individual written description of a chosen solving strategy (30 min.)
c) the division of the class into two groups discussing the task (30 min.)
d) the proof of a strategic processing given by the competitive groups (30
min.)
Pupils from higher secondary school solved the same problem like the
pupils from middle school in the same way and within the same time limit.
Individual works were analyzed (a-priori analysis), the identification of
parameters was carried out and those were subsequently used as a basis for
the characteristics of pupils’ answers. It enabled to do a quantitative analysis
of the answers, to establish an implicative graph (graph of functionality), a
hierarchical diagram, a diagram of similarities and also an analysis of data.
The analyses, graphs and diagrams (or trees) were part of the evaluation of
each experiment together with conclusions.

2.5 The first experiment and its analysis by CHIC

The first statistical survey was made by using a sample of 88 pupils attending
the third and fourth year of study of secondary school in Palermo (Sicily). The
students worked in pairs for the part relating to interviews and individually
for the production of solution protocols related to the proposed conjecture.
The variables used for the a-priori analysis were 15 and they were settled in
the following manner:
1. He/she verifies the conjecture by natural numbers taken at random. (N-
random)
2. He/she sums two prime numbers at random and checks if the result is an
even number. (Pr-random)
3. He/she factorizes an even number and sums its factors, trying to obtain
two primes. (Factor)
4. Golbach’s method 1. He/she considers odd prime numbers lesser than an
even number, summing each of them with successive primes. (Gold1)
5. Golbach?s method 2 (letter to Euler). He/she writes an even number as a
sum of more units, combining these in order to get two primes. (Gold2)
6. Cantor?s method. Given an even number 2n, by subtracting from it the
prime numbers x ≤ 2n one by one, by a table of primes one tempts if
the obtained difference 2n − x is a prime. If it is, then 2n is a sum of two
primes.(Cant)
 Modelling by Statistic in Research of Mathematics Education 259

7. The strategy for Cantor’s method. He/she considers the primes lower then
the given even number and calculates the difference between the given
number and each of primes. (S-Cant)
8. Euler. He/she is uneasy to prove the conjecture because one has to con-
sider the additive properties of numbers. (Euler)
9. Chen Jing-run?s method (1966). He/she expresses an even number as a
sum of a prime and of a number which is the product of two primes.(Chen)
10. He/she subtracts a prime number from an any even number (lower then
the given even number) and he/she ascertains if he/she obtains a prime,
so the condition is verified. (Spa-pr)
11. He/she looks for a counter-example which invalidates the statement of the
conjecture. (C-exam)
12. He/she considers the final digits of a prime to ascertain the truth of the
statement.(Cifre)
13. He/she thinks that a verification of the statement by some numerical
examples needs to prove the statement. (V-prova)
14. He/she does not argue anything for the second question. (Nulla)
15. He/she thinks the conjecture is a postulate. (Post)

2.6 The implicative graph

The analysis of the implicative graph shows, with percentages of 90%, 95%
and 99%, that pupils’ choice to follow some of the strategies is strictly linked
to a relevant strategy, namely Gold 1, or the one according which the pupil
considers odd prime numbers summing each of them with successive primes.
Hence the basis of pupils’ behaviours is the sequential thinking.

2.7 The factorial analysis by S.P.S.S.

The graph shows that a part of pupils is inclined either to proceed by a

sequential manner or by preferring a method based on a random choice. On the
other hand, the second component shows that the real strong characterization
of most pupils is Gold1-Chen which is nearer to the intersection of the two
components. So, this is the winning strategy among pupils to pass from an
argumentation to a possible demonstration of the conjecture. This is a kind
of a photo of the more frequent approaches to the conjecture by students.

2.8 Supplementary variables and pupils’ profiles

A further step was made by introducing three supplementary variables to get

other informations about the obtained data. They were the following ones:
a) Abdut: this is the pupil proceeding by abduction, which indicates (Peirce)
the first moment of an inductive process, when a pupil chooses a hypoth-
esis by which he can explain determined empirical facts. On the base of
260 E. Malisani et al.

Fig. 2. The implicative graph 90%.

Component 2

Component 1

Fig. 3. Factorial analysis variables

 Modelling by Statistic in Research of Mathematics Education 261

such a definition, the pupil named Abdut is who observes how Goldbach’s
conjecture to be verified in a large number of cases, so he supposes it
is also valid for any very large even number, and this fact leads him to
the final thesis, namely the conjecture to be valid for every even natural
number.
b) Intuitionist: this is the pupil using the N-random and Euler strategies
in common with Abdut, but thinking that the demonstration of the con-
jecture can be deduced by a simple numerical evidence, because he is
convinced that what happens for the elements of a small finite set of val-
ues can be generalized to the infinite set which the small set belongs to.
So, he uses the V-prova strategy. In short, in an inductive argumentation
used by the intuitionist pupil the statement is deduced as a generic case
after specific cases.
c) Ipoded: this is just the pupil using a deductive argumentation which can
be directly transposed into a deductive demonstration.
With these new additional variables a transposed matrix (changing rows by
columns) was made by Excel and interpreted by CHIC. The more interesting
results are displayed in figure 4.

Fig. 4. The implicative graph with supplementary variables

2.9 The implicative graph

It is evident that the three profiles corresponding to the additional variables

are significative as much as they catalyze the outlines of reasoning of the
pupils. The supplementary variables play in this case the same role played by
262 E. Malisani et al.

an equivalence relation when we put it into a set. In fact, what the relation does
in this case? It orders the elements of the set, so the supplementary variables,
in this case, give much more order to the data. They get the interpretation of
data more effective. Really, they begin attractors for pupils’ behaviours.

2.10 Factorial Analysis

Fig. 5. Factorial analysis with supplementary variables. Component 1 is refered to

abscissae,Component 2 is refered to ordinata

From the viewpoint of the horizontal component the variable Intuitionist

characterizes it weakly, while the variables Abdut and Ipoded with a lot of other
variables characterize it much more. On the other hand, this is a paradigmatic
situation which has its historical counterpart in the attempts made along
centuries by different mathematicians facing the conjecture. So, Abdut and
Ipoded profiles are winners, while the intuitive method of approach is less
productive. This characteristic situation is stationary also when one observes
the graph from the viewpoint of the second component. This means that in
any way Abdut and Ipoded methods are more interesting for pupils.

2.11 Some final observations

The experimentation about Goldbach’s conjecture has pointed up that in

general most pupils, while facing an unsolved historical conjecture (without
knowing it is yet unsolved), start at once with an empirical verification of it
which can support their intuition, but after they distinguish themselves along
three different solving typologies:
 Modelling by Statistic in Research of Mathematics Education 263

1. a part of pupils bites off more than one can chew with the following
conclusion: since the conjecture is true for all of these particular cases,
then it has to be true anyway. These are pupils who have a strong faith in
their convictions, but who do not know clearly enough how to pass from
an argumentation to a demonstration, by using the achieved data.
2. a part of pupils proceeds at the same time by an empirical verification and
by an attempt of argumentation and demonstration ending to a mental
statement. They try to clear a following hurdle: how can I deduce a gen-
eral statement from the empirical evidence? These are pupils who before
making any generalization want to be sure of the made steps, therefore
they tread carefully.
3. few pupils, after a short empirical verification, look at once for a formal-
ization of their argumentations, but if they are not able to do that, they
are not diffident about claiming they are in front of something which is
undemonstrable. These pupils have a high consideration for their mental
processes therefore they think that if they are not able to demonstrate
anything, then it has to be undemonstrable anyway.
By this experimentation we argue that the argumentation favoured by
pupils facing a historical conjecture like Goldbach’s is the abductive one. Some
questions arise from the results which would be advanced by other experimen-
tations:
1. Is this result generalizable?
2. To what extent is it generalizable?
But the fundamental kernel of this experimentation about the interplay
between history of mathematics and mathematics education is that such re-
sults could not be pointed out if the a-priori analysis had not been made by
the historical-epistemological remarks which have inspired it.

3 The Statistic Implicative analysis

and the correspondence factor analysis in a research
in Mathematics Education: unknown or “thing which
is varying”
3.1 Introduction

Modelling through statistic argumentation gives to the research in mathemat-

ics education a greater possibility of transferability of the experience. However,
the statistic argumentation would not have any weight without an accurate
theoretical reflection from the viewpoint of the didactic and the epistemology
of the mathematical contents [11].
In the a-priori analysis of a didactic situation it is necessary to consider the
epistemological representations, the historical-epistemological representations
264 E. Malisani et al.

and the hypothetical behaviours, correct and not correct. Besides the a-priori
analysis allows us to individualize the variables of the situation-problem and
the hypotheses of research. These hypotheses can be falsified through the
statistic analysis and/or the qualitative analysis of the data.
In the last decade two statistic methods have been very used: the im-
plicative analysis (ASI) of Regis Gras [13, 14] and the correspondence factor
analysis (CFA). The implicative analysis is a powerful tool. It allows a clear
visualization of the relations of similarity and implication among variables or
classes of variables of the situation-problem, through the graphs elaborated
by the software CHIC. The correspondence factor analysis represents geomet-
rically, in a multi-dimensional space, a distribution of two set: the individuals
and the variables of the situation [13]. Since it allows an analysis on small
samples in the field of the not parametric Statistics, it contributes to inter-
pret meaningfully the didactic phenomena. In the last decade the aim of some
studies was to improve the tool in the field of the didactic research and, chiefly,
to create some models ad hoc [35].
This paper is a contribution to the studies on the application of the Statis-
tic Implicative Analysis (SIA) and the Correspondence Factor Analysis (CFA)
in different fields, particularly, in Mathematics Education. This research puts
in evidence the relations between the Implicative Analysis and the Factorial
Analysis to falsify hypotheses of research in mathematics education. We want
to analyze too the type of information obtained by the application of the two
statistic methods12 .

3.2 A condensed theoretical framework

There are a lot of studies on the obstacles which pupils meet during the
passage from the arithmetic to the algebraic thought. Some of them reveal
that the introduction of the concept of variable represents the critical point
of transition [25, 39, 40].
This is a complex concept because it is applied with different meanings in
different situations. Its management depends precisely on the particular way
of its use into the activity of problem-solving.
The notion of variable could take on a plurality of conceptions: general-
ized number (it appears in the generalizations and in the general methods);
unknown (its value could be calculated by considering the restrictions about
the field of existence of the solutions of a problem); “in functional relation”
(relation of variation with other variables); sign totally arbitrary (it appears
in the study of the structures); register of memory (in informatics) [38].
In Malisani and Marino [23] and Malisani and Spagnolo [24] we observed
that the pupils spontaneously evoke the different conceptions of variable as:
numerical value, unknown, “thing which is varying”, also in absence of an
adequate mastery of the algebraic language.
12
This part of the paper is based on [22]
 Modelling by Statistic in Research of Mathematics Education 265

It is possible that many difficulties in the study of algebra derive from the
inadequate construction of the concept of variable [8]. An opportune approach
to this concept should consider its principal conceptions, the existing inter-
relationships between them and the possibility to pass from one to the other
with flexibility, in relation to the exigencies of the problem to be solved.
The historical analysis emphasizes that the notions of unknown and the
one of variable as “thing which is varying” have a totally different origin and
evolution. Even if both the concepts deal with numbers, their processes of
conceptualizations seem to be entirely different [26].
In Malisani [20, 21] we studied the relational-functional aspect of the vari-
able in problem-solving, considering the semiotic contexts of algebra and ana-
lytical geometry. We showed that there is a certain interference of the concep-
tion of unknown on the functional aspect, in the context of a situation-problem
and in absence of visual representative registers. We also demonstrated that
the students find some difficulties to interpret the concept of variable in the
process of translation from the algebraic language into natural one.
This paper belongs of the statistic analysis of that experimentation, by
which we wanted to verify if the conception of variable as “thing which is
varying” is evoked, when the notion of unknown prevails in the context of a
situation-problem.
To carry out this research we chose the linear equation in two variables
for two reasons: firstly, because it represents a nodal point from which the
students derive the conceptions of the letters as unknowns or “things which
are varying”. Secondly, this kind of equation is well known by the pupils,
because they studied it from different viewpoints: linear function, equation of
a straight line and component of the linear systems.

3.3 Methodology of the research

One hundred eleven students — aged 16–18 — of the Experimental High

School of Ribera (AG, Italy) participated to the research.
The questionnaire presentes four questions, but in this paper we introduce
only the resolution of the first problem which is the following:
Charles and Lucy win the total sum of a Euro 300 in the lottery. We know
that Charles wins the triple of the betted money, while Lucy wins the quadruple
of her own.
1. Determine the sums of money that Charles and Lucy have betted. Com-
ment on the procedure that you have followed.
2. How many possible solutions are there? Give reasons for your answer.
In this problem the variable takes on the relational-functional aspect in the
context of a concrete situation-problem. We also asked the pupils to think
over the solution set. With this question we wanted to analyze the solving
strategies used and if the unknown’s notion interferes with the interpretation
of the functional viewpoint.
266 E. Malisani et al.

We carried out a-priori analysis of the problem. The aim was to determine
all the possible strategies that the pupils could use. Some errors that students
could possibly make in the application of these strategies were also identified.
The pupils worked individually, we did not allow them consulting books
or notes. The where given time was sixty minutes.
In a table we filled in with a double input “pupils/strategy”, and we in-
dicated for every pupil the strategies he used by the value 1 and those he
didn’t apply by the value 0. The data were analyzed in a quantitative way,
by using the statistic implicative analysis of Regis Gras [13, 14], the software
CHIC 2000 and the factorial statistical survey S.P.S.S. (Statistical Package
for Social Sciences).

3.4 The a-priori analysis

We determined the principal experimental variables from an a-priori analysis.

They were the following ones:
AL1: The pupil answers the question.
AL2: He/she shows a procedure in the natural language.
AL3: He/she shows a procedure by trial and errors in natural language
and/or in a half-formalized language.
AL4: He/she adds a datum.
AL4.1: He/she adds a datum, but he/she considers that the winnings are
divided in half.
AL4.2: He/she adds a datum, but he/she considers that the winnings of
the two teenagers are equal to Euro 300.
AL4.3: He/she adds a datum, but he/she considers that the bets are equal.
AL5: He/she translates the problem into a first degree equation of two
unknowns.
AL7: He/she translates the problem into a first degree equation with two
unknowns and he/she uses the algebraic method of “substitution into the same
equation” 13 .
AL9: He/she abandons the pseudo-algebraic procedure and he/she tries
another method.
AL11: He/she considers, in an explicit or implicit way, that the problem
represents a functional relation.
AL13: He/she makes some errors in the resolution of the equation and
he/she finds (or he/she tries to find) the only solution.
AL14: He/she considers that a relation of proportionality exists between
x and y.

13
We called “procedure of substitution into the same equation” the incorrect method
where he/she writes one variable in function of the other. Then he/she replace this
variable in the original equation, and thus he/she obtain an identity. In short, the
pupil applies the method of substitution used to solve the systems of equations
to a single equation.
 Modelling by Statistic in Research of Mathematics Education 267

ALb1: The pupil calculates the solution set.

ALb2: He/she shows a particular solution verifying the equation.
ALb3: He/she shows several solutions verifying the equation.
ALb4: He/she considers the infinite solutions expressly.
ALb5: He/she explicitly considers the data are insufficient to determine
only one solution.
ALb6: He/she considers multiple solutions (it includes ALb4 and ALb5).

The hypothesis

If the conception of variable as un- then the relational-functional aspect is

known prevails in the context of a not evoke
situation-problem
AL4, ALb2 ∼AL3, ∼AL11, ∼AL14, ∼ALb3,
∼ALb4, ∼ALb5, ∼ALb6

Table 2. Hypothesis and experimental variables

The conception of variable as an unknown is highlighted by the experi-

mental variable AL4 “he/she adds a datum”. Precisely, “adding a datum” is
equivalent to introducing a new equation and thus to forming a system of two
linear equations with the equation of the problem or part of it. The solution
of the system is a “particular solution verifying the equation of the problem”
(ALb2).
The relational-functional aspect of the variable is evoked when the pupil
exhibits a “procedure by trial and errors” (AL3), through which he recalls the
notion of dependence among the variables. In this way, the pupil “consid-
ers implicitly or expressly that the problem represents a functional relation”
(AL11) or he manifests, incorrectly, that “a relation of direct proportionality
exists among the variables” (AL14). Therefore the pupil “shows some solutions
veryfing the equation” (ALb3) or “he considers that the problem has multiple
solutions” (ALb6)14 Accordingly “not to evoke the relational-functional aspect
of the variable” is equivalent to the negation of the experimental variables
above described: ∼AL3, ∼AL11, ∼AL14, ∼ALb3, ∼ALb4, ∼ALb5, ∼ALb6.

14
In this study we prefer to use the term “multiple solutions” rather than the one
of “infinite solutions”, because we have not considered the possible connotations
of the word “infinite”. However, we defined two experimental variables ALb4 to
take into account the cases in which the pupil explicitly considers the existence
of infinite solutions.
268 E. Malisani et al.

3.5 Statistic Implicative Analysis (ASI)

Implicative graph

AL14 AL6
AL9 AL13
AL4.2 AL4.3 ALb5 AL7
ALb4 AL5
AL3 ALb3
ALb6 AL11
AL2
AL4.1
ALb2
AL4

ALb1
AL1
99 95 90 85

Fig. 6. Implicative graph

The implicative graph on Figure 6 (carried out with the software CHIC
2000) shows three well defined groups of experimental variables with statis-
tic percentages of 95% and 99%. They are pointed out by the cloud on the
left (cloud L), the cloud in the center of the figure (cloud C) and the cloud
on the right (cloud R) (the grey cloud (cloud I) around AL11 indicates the
intersection between cloud C and cloud R).
The three groups are directly or indirectly connected with the variable
ALb1 “the pupil calculates the solution set” and AL1 “the pupil answers to the
question”. Every group corresponds to a different kind of strategy used by the
students:
• Procedure in natural language (cloud L): the pupil adds a datum
considering that the winnings are equal (generally dividing Euro 300 in
half) or that the bets are equal15 . In this way, the student transforms the
question into a typical arithmetic problem and he resolves it finding only
a particular solution verifying the equation. This result is confirmed by
15
The experimental variable AL4 “he/she adds a datum” considers two possibilities:
equal winnings or equal bets (AL4.3). The first case takes into account the two
other alternatives: the winnings are divided in half (AL4.1) or both the teenagers
win Euro 300 (AL4.2).
“To add a datum” is equivalent to introduce a new equation and to forme (with
the equation of the problem 3x+4y = 300 or part of it) a system of two equations
into two unknowns. Therefore a system corresponds to each case.
“The winning of Euro 300 are divided in half ” (AL4.1): it is equivalent to the
system 3x + 4y = 300, 3x − 4y = 150
 Modelling by Statistic in Research of Mathematics Education 269

the implicative links among the experimental variables AL2, ALb2 and
AL4 (with its variations AL4.1, AL4.2 and AL4.3). The procedure in the
natural language is the most used by the pupils (Cfr. Table of frequencies
in the Appendix ), and it leads to the single solution. So the predominant
conception of variable is that of unknown.

• Method by trials and errors in natural language or in half-

formalized language (cloud C): the pupil generally assigns several values
to a variable (e.g. Charles’ bet) and he finds the corresponding values in
the other variable (Lucy’s bet). In this way, the student shows some so-
lutions verifying the equations and/or he considers that it has multiple
solutions. That is, he generally considers in an implicit way that the prob-
lem represents a functional relation. This result result is obtained by the
implicative links between the variables AL3, ALb3, AL11 and ALb6. This
method leads to many solutions, allows evoking the dependence between
the variables, but a strong conception of the relational-functional aspect
does not appear yet.

• Pseudo-algebraic strategy (cloud R): the pupil translates the text of

the problem into an equation of first degree with two unknowns and applies
the method of “substitution into the same equation”, that is the incorrect
procedure where he/she writes one variable in function of the other. Then
he/she replaces this variable it the original equation and thus he/she obtain
an identity. Since the pupil does not succeed in interpreting the identity,
either he/she changes his/her resolving procedure abandoning the pseudo-
algebraic one or he/she resumes the resolution of the equation and he/she
makes some errors trying to find only one solution. This result is deduced
by the implicative links among the experimental variables AL9, AL13, AL7
and AL5.
It is interesting to observe that, if the pupil abandons this strategy, then
he/she considers, in an implicit or explicit way, that the problem repre-
sents a functional relation. This result is confirmed by the implicative link
between the experimental variables AL9 and AL11. This link allows the con-
nection between the two procedures: by trials and errors and pseudo-algebraic
(cloud I).
However, the pseudo-algebraic strategy is rarely used and it leads to the
correct solution of the problem only in some cases.

Falsification of the hypothesis

We are considering:
“The winnings are equal to Euro 300 for both the teenagers” (AL4.2): it cor-
responds to the system 3x = 300, 4y = 300
“The bets are equal” (AL4.3): is equivalent to the system
3x + 4y = 300, x = y
270 E. Malisani et al.

p: in the context of a situation-problem, the conception of variable as unknown

(experimental variables AL4 and ALb2) prevails;
q: the relational-functional aspect is evoked (experimental variables AL3,
AL11, AL14, ALb3, ALb4, ALb5 and ALb6).
The hypothesis 1 is equivalent to:
p →∼ q that, from the logical viewpoint, is equivalent to ∼ (p∧ ∼ (∼ q))
or ∼ (p ∧ q).
Therefore, to falsify this hypothesis it is sufficient to demonstrate the
empty intersection between the experimental variables of p and q, in other
words:
p corresponds to the procedure in the natural language which the conception
of variable as unknown prevails into;
q is equivalent to the method by trials and errors which the relational-
functional aspect of the variable predominates into.
From the implicative graph we deduce that the sets of experimental variables,
corresponding to p (cloud L) and to q (cloud C), are disjoined. This result
allows us to falsify the formulated hypothesis.

Profile of the pupils

From the previous analysis an important aspect emerges: the existence of

a certain correspondence between the solving strategies used by the pupils
and the conceptions of variable as unknown and “thing which is varying”. To
examine carefully these results we apply a particularly effective methodology
by introducing some supplementary variables in the “students” component.
In other words, the researcher defines a profile of student that satisfies some
characteristics he considers important. In the specific case of this analysis we
define five profiles:
1. NAT: this profile corresponds to the pupil performing a procedure in
the natural language. Then he/she adds a datum considering that the
winnings are equal (generally dividing it in half) or that the bets are
equal and he/she resolves the problem finding only a particular solution
verifying the equation. This profile is characterized by the presence of the
followings experimental variables: AL2, AL4 (AL4.1, AL4.2 or AL4.3) and
ALb2.
2. FUNZ: it corresponds to the pupil applying a strategy by trials and errors
in natural language and/or in half-formalized language. He/she generally
assigns several values to a variable and he/she finds the corresponding
values in the other variable. So the student shows some solutions verify-
ing the equations and/or he/she considers that it has multiple solutions.
The experimental variables describing this profile are: AL3, AL11, ALb3,
ALb4, ALb5 and ALb6.
 Modelling by Statistic in Research of Mathematics Education 271

3. PALG1: it corresponds to the student who uses the pseudo-algebraic pro-

cedure. He/she translates the text of the problem into a linear equation
with two unknowns and he/she applies the method of “substitution into
the same equation”. When the student reaches the identity he/she does
not succeed in interpreting it. Then he/she resumes the resolution of the
equation, he/she makes some errors of syntactic kind trying to find only
one solution. This profile is characterized by the presence of the experi-
mental variables: AL5, AL7, AL13 and ALb2.
4. PALG2: it is a variation of the profile PALG1. In this case, when the pupil
arrives to the identity he/she changes the solving procedure abandoning
the pseudo-algebraic one. The experimental variables which describe this
profile are the followings: AL3, AL5, AL7, AL9, AL11 and ALb3.
5. ALG: it corresponds to the pupil who applies an algebraic procedure.
He/she translates the problem into an equation of first degree with two
unknowns, he/she considers, in an implicit or explicit way, that it repre-
sents a functional relation and so that it is verified by multiple solutions.
The experimental variables of this profile are the followings: AL5, AL11,
ALb4 and ALb6.

3.6 The correspondence factor analysis (CFA)

Fig. 7. Factorial analysis with supplementary variables

The graph shows that the first component (horizontal axis) is strongly
characterized by the pair of supplementary variables: NAT and PALG1.
The profiles ALG, PALG2 and FUNZ form a cloud that strongly character-
izes the vertical component. The supplementary variable PALG2 is very near
to FUNZ, because the student who abandons the pseudo-algebraic procedure
generally adopts the profile described in FUNZ.
272 E. Malisani et al.

The winning strategies are precisely those described in the profiles ALG,
PALG2 and FUNZ which lead to multiple solutions, while NAT and PALG1
lead to the oneness of the solution. This finds a strong correspondence with
the different conceptions of “variable”. Therefore, the horizontal axis represents
the conception of variable as unknown, the vertical axis, instead, reproduces
its relational-functional aspect. These results allow us to falsify the hypothesis
again.

3.7 Conclusions

The implicative graph shows the solving strategies applied by the students to
solve the problem:
1. procedure in natural language: it is the most used by the pupils and it
leads to the single solution. The predominant conception of variable is
that of unknown.
2. methods by trials and errors in natural language and/or in half-formalized
language (generally arithmetic): it gets to several solutions. The depen-
dence of the variables is evoked, but a strong conception of the relational-
functional aspect does not appear yet.
3. pseudo-algebraic strategy: it is little used by the pupils and it leads to the
correct solution of the problem only in some cases.
To examine carefully these results we introduced some supplementary vari-
ables in the “students” component. These profiles represent the supplementary
individuals putting out the fundamental characteristics of the a-priori analy-
sis. They are displayed in Table 3.

SUPPLEMENTARY STRATEGIES
VARIABLES
NAT in natural language
FUNZ by trials and errors in natural language and/or
in half-formalized language
PALG1 pseudo-algebraic + resolution of the equation with
some errors of syntactic kind
PALG2 pseudo-algebraic + other strategy
ALG algebraic

Table 3. Correspondence between supplementary variables and strategies

The hierarchic tree shows that the profile NAT is the most meaningful
because it represents the strategy the pupil used most. We observe a small
set of pupils connected to this group. They followed the procedure described
in NAT; but, afterwards, they effected the passage from the single solution to
multiple solutions [22, pp. 96].
 Modelling by Statistic in Research of Mathematics Education 273

From the factorial analysis we observe that the horizontal component is

characterized by the profiles NAT and PALG1, while the vertical compo-
nent by FUNZ, PALG2 and ALG. Therefore we note a strong correspondence
among the principal components and the conceptions of variable: the hori-
zontal component represents the notion of unknown, while the vertical one
denotes the aspect of “thing which is varying”.
The obtained results, by the implicative analysis and the factorial analysis,
allow us to falsificate the formulated hypothesis, namely: “If the conception of
variable as unknown prevails in the context of a problematic situation, then
the relational-functional aspect is not evoked”.
It is interesting to observe that, in some cases, we verify the passage from
the single solution to multiple solutions in the linear equation, even if the
notion of unknown prevails. From here an important question emerges: “This
passage coincides (or not) with the passage from the conception of unknown
to that relational-functional one”? In this experimentation we have not found
the answer. To study carefully this matter we carried out a new experimental
research of qualitative kind, submitting the same questionnaire to pairs of
pupils.
This research puts in evidence the relations between the Statistic Implica-
tive Analysis (ASI) and the Correspondence Factor Analysis (CFA) to falsify
hypotheses of didactic research in mathematics. The implicative analysis puts
in evidence the strategies of the students, the factorial analysis puts in contrast
the correspondence between the principal components and the conceptions on
the variable.

References
1. R. Agrawal, T. Imielinski, and A.N. Swami. Mining association rules between
sets of items in large databases. In P. Buneman and S. Jajodia, editors, ACM
SIGMOD International Conference on Management of Data, pages 207–216,
1993.
2. F. Arzarello, L. Bazzini, and G. Chiappini. L’algebra come strumento di pen-
siero. analisi teorica e considerazioni didattiche. Progetto Strategico CNR-TID,
(6), 1994.
3. Ch. Bastin, J.P. Benzecri, Ch. Bourgarit, and P. Cazes. Pratique de l’Analyse
des Données, volume 1–2. Dunod, 1980.
4. A. Bodin. Improving the diagnostic and didactic meaningfulness of mathematics
assessment in france. In Annual Meeting of the American Educational Research.
Association AERA, New York, 1996.
5. G. Brousseau. Theory of didactical situations in mathematics. Kluwer Aca-
demic Publishers, 1997. Edited and translated by N. Balacheff, M. Cooper,
R. Sutherland and V. Warfield.
6. G. Brousseau. Théorie des situations didactiques. Didactique des mathématiques
1970–1990. La pensée sauvage, 1998. Textes rassemblés.
274 E. Malisani et al.

7. E. La Casta and G. Brousseau. Méthodes d’analyses statistiques multidimen-

sionnelles en didactique des mathématiques, chapter Utilisation de la contin-
gence par l’analyse factorielle. Traitement d’un cas: Le graphique, pages 53–90.
ARDM, Rennes, 1995.
8. I. Chiarugi, G. Fracassina, F. Furinghetti, and D. Paola. Parametri, variabili
e altro: un ripensamento su come questi concetti sono presentati in classe.
L’insegnamento della Matematica e delle Scienze integrate, 18B(1):34–50, 1995.
9. R. Couturier. Traitement de l’analyse statistique implicative dans chic. In Actes
des Journées sur la “Fouille dans les données par la méthode d’analyse implica-
tive”, 2001.
10. B. Escofier and J. Pages. Analyses factorielles simples et multiples (objectifs,
méthodes et interprétation). Dunod, Paris, 1990.
11. F.Spagnolo. Insegnare le matematiche nella scuola secondaria. La Nuova Italia,
Firenze, 1998.
12. R. Gras. L’implication statistique (Nouvelle méthode exploratoire de données).
La Pensée Sauvage, Grenoble, 1996.
13. R. Gras. Metodologia di analisi di indagine. Quaderni di Ricerca Didattica, 7:99–
109, 1997.
14. R. Gras. I fondamenti dell’analisi statistica implicativa. Quaderni di Ricerca
Didattica, 9:189–209, 2000. Text available at: http://dipmat.math.unipa.it/
grim/quaderno9.htm.
15. R. Gras, R. Couturier, F. Guillet, and F. Spagnolo. Extraction de règles en
incertain par la méthode statistique implicative. In Comptes rendus des 12e
Rencontres de la Société Francophone de Classification, pages 148–151, 2005.
16. R. Gras, E. Diday, P. Kuntz, and R. Couturier. Variables sur intervalles et
variables-intervalles en analyse implicative. In Actes du 8e Congrès de la Société
Francophone de Classification, pages 166–173, 2001.
17. J.B. Lagrange. Analyse implicative d’un ensemble de variables numériques;
application au traitement d’un questionnaire aux réponses modales ordonnées.
Revue de Statistiques Appliquées, 46(1):71–93, 1998.
18. I.C. Lerman. Classification et analyse ordinale des données. Dunod, 1981.
19. I.C. Lerman, R. Gras, and H. Rostam. Elaboration et évaluation d’un indice
d’implication pour des données binaires, i et ii. Mathématiques et Sciences
Humaines, 74 and 75:5–35 and 5–47, 1981.
20. E. Malisani. The notion of variable in semiotic contexts different. In Proc. of
the Int. Conf. “The Humanistic Renaissance in Mathematics Education”, pages
245–249, University of Palermo-Italy, 2002. Text available at: http://dipmat.
math.unipa.it/~grim/21project.htm.
21. E. Malisani. The notion of variable: some meaningful aspects of algebraic lan-
guage. In A. Gagatsis, F. Spagnolo, G. Makrides, and V. Farmaki, editors, Proc.
of the 4th Mediterranean Conf. on Mathematics Education (MEDCONF 2005),
volume 2, pages 397–406, University of Palermo-Italy, 2005.
22. E. Malisani. The concept of variable in the passage from the arithmetic language
to the algebraic language in different semiotic contexts. PhD thesis, Palermo,
Italy, 2006.
23. E. Malisani and T. Marino. Il quadrato magico: dal linguaggio aritmetico al
linguaggio algebrico. Quaderni di Ricerca in Didattica, 10, 2002. Text available
at: http://dipmat.math.unipa.it/~grim/quaderno10.htm.
24. E. Malisani and F. Spagnolo. Difficulty and obstacles with the concept of vari-
able. In Proc. of CIEAEM 57, pages 226–231, Piazza Armerina-Italy, 2005.
 Modelling by Statistic in Research of Mathematics Education 275

25. M. Matz. Intelligent Tutoring Systems, chapter Towards a Process Model for
High School Algebra Errors, pages 25–50. Academic Press, London, 1982.
26. L. Radford. Approaches to Algebra. Perspectives for Research and Teaching,
chapter The roles of geometry and arithmetic in the development of algebra:
historical remarks form a didactic perspective, pages 39–53. Kluwer, 1996.
27. A. Scimone. Following goldbach’s tracks. In Proc. of the Int. Conf. “The Hu-
manistic Renaissance in Mathematics Education”, University of Palermo-Italy,
2002. Text available at: http://dipmat.math.unipa.it/~grim/21project.htm.
28. A. Scimone. La congettura di goldbach tra storia e sperimentazione didattica.
Quaderni di Ricerca in Didattica, 10:1–37, 2002. Text available at: http://
dipmat.math.unipa.it/~grim/quaderno10.htm.
29. A. Scimone. Conceptions of pupils about an open historical question: Goldbach’s
conjecture. The improvement of Mathematical Education from a historical view-
point. PhD thesis, Palermo, Italy, 2003. published on Quaderni di Ricerca
in Didattica 12, Text available at: http://dipmat.math.unipa.it/~grim/
tesi/_it.htm.
30. A. Scimone. An educational experimentation on goldbach’s conjecture. In Proc.
CERME 3, Group 4, pages 1–10, Bellaria-Italy, 2003.
31. A. Scimone. How much can the history of mathematics help mathematics educa-
tion? an interplay via goldbach’s conjecture. In Zbornik, Bratislavskehoseminara
z teorie vyucovania matematiky, pages 89–101, Bratislava, 2003.
32. F. Spagnolo. Obstacles Epistémologiques: Le Postulat d’Eudoxe-Archimede. PhD
thesis, Universiy of Bordeaux I, 1995.
33. F. Spagnolo. L’analisi a priori e l’indice di implicazione di regis gras. Quaderni
di Ricerca in Didattica, 7:110–117, 1997. Text available at: http://dipmat.
math.unipa.it/~grim/quaderno7.htm.
34. F. Spagnolo. A theoretical-experimental model for research of epistemological
obstacles. In Int. Conf. on Mathematics Education into the 21st Century, 1999.
Text available at: http://dipmat.math.unipa.it/~grim/model.pdf.
35. F. Spagnolo. L’analisi quantitativa e qualitativa dei dati sperimentali. Quaderni
di Ricerca in Didattica, 10, Supplemento, 2002. Text available at: http://
dipmat.math.unipa.it/~grim/quaderno10.htm.
36. F. Spagnolo. La modélisation dans la recherche en didactiques des mathéma-
tiques: les obstacles épistémologiques. In Recherches en Didactiques des Math-
ématiques, volume 26. La Pensée Sauvage, Grenoble, 2006.
37. F. Spagnolo and R. Gras. Fuzzy implication through statistic implication:
a new approach in zadeh’s framework. In S. Dick, L. Kurgan, W. Pedrycz,
and M. Reformat, editors, Proc. of Annual Meeting of the North American
Fuzzy Information Processing Society (NAFIPS 2004), volume 1, pages 425–429,
Banff,Canada, 2004.
38. Z. Usiskin. Conceptions of school algebra and uses of variables. In A.F. Coxford
and A.P. Shulte, editors, The ideas of Algebra, pages 8–19. NCTM, Reston-Va,
1988.
39. S. Wagner. An analytical framework for mathematical variables. In Proc. of the
Fifth PME Conference, pages 165–170, Grenoble, France, 1981.
40. S. Wagner. What are these things called variables. Mathematics Teacher,
76(7):474–479, 1983.
276 E. Malisani et al.

Appendix: Table of frequencies

Variable Absolute frequency Relative frequency Percentage Rest

AL1 106.00 0.95 95 0.21
AL2 44.00 0.40 40 0.49
AL3 34.00 0.31 31 0.46
AL4 71.00 0.64 64 0.48
AL4.1 53.00 0.48 48 0.50
AL4.2 11.00 0.10 10 0.30
AL4.3 13.00 0.12 12 0.32
AL5 27.00 0.24 24 0.43
AL6 4.00 0.04 4 0.19
AL7 13.00 0.12 12 0.32
AL9 9.00 0.08 8 0.27
AL11 34.00 0.31 31 0.46
AL13 8.00 0.07 7 0.26
AL14 3.00 0.03 3 0.16
ALb1 99.00 0.89 89 0.31
ALb2 63.00 0.57 57 0.50
ALb3 33.00 0.30 30 0.46
ALb4 25.00 0.23 23 0.42
ALb5 13.00 0.12 12 0.32
ALb6 36.00 0.32 32 0.47
Didactics of Mathematics and Implicative
Statistical Analysis

Dominique Lahanier-Reuter

Université Charles-de-Gaulle,
Equipe THEODILE E.A. 1764
59653 Villeneuve d’Ascq, France
[email protected]

Summary. People working in Didactics of Mathematics have constantly regarded

statistical implicative analysis as a profitable and heuristic method of data analysis.
First we intend to show the reasons for this interest: implicative links that S.I.A.
has pointed out may be interpreted as rules and regulations connecting actions,
discourses, . . . , or as a group’s characteristics. We develop some examples showing
how S.I.A. can be used and what special research results it can provide. We insist
upon some points that may be interesting methodologically to focus on: asymmetric
links, nodes and separate implicative ways.

Key words: Mathematical didactics, rules, regulations, school subjetc, geometric

task, geometric skill

Theorizing the relations between these two fields of research, didactics

of mathematics and Statistical Implicative Analysis, in terms of producers of
models and techniques on one hand and in the field of application on the other
is without doubt too reductive, from a historical point of view. In fact the still
very brief history of the emergence of these two scientific domains shows us
some more complex connections: the coincidence of the time of emergence
and recognition, that of their geographical and institutional situations- some
universities and associations of researchers in France —and finally, and above
all, the presence of common actors — Regis Gras particularly [4–6]— imply a
dynamic which is specific to these relations. Thus the implicative analysis of
data has been able to be identified as a preferred method of analysis in math-
ematical didactics and reciprocally in some way, some problems of didactics
have been able to raise questions in implicative analysis.
First we seek to clarify the reasons why we see it as a fruitful cooperation,
in seeking to understand how the clarification of rules (or quasi-rules) that
implicative analysis allows proves to be a valuable and pertinent tool for
mathematical didacticians. Then we will explain two of the main problems
in which these rules have meaning in mathematical didactics: that of the

D. Lahanier-Reuter: Didactics of Mathematics and Implicative Statistical Analysis, Studies in

Computational Intelligence (SCI) 127, 277–298 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
278 D. Lahanier-Reuter

regulation of observable behaviours of students in a situation and that of

controls, in this instance understood as effects of teaching devices.

1 Rules and Regulations

The implicative analysis of data allows us to show rules (or quasi rules) that
structure a set of data from calculations of the co-occurrences of some modal-
ities of variables. These rules can be generically represented by an expression
of the type “if A then B”. They are consequentially hierarchical, that means
that they operate an asymmetry between the regulated modalities. In this,
the implicative analysis of data (henceforth S.I.A) is distinguished from other
modes of statistical analysis that, if they are equally based on calculations
of co-occurrences of modalities of variables, only exhibit symmetrical rules
that therefore do not discriminate between the variables studied. Studying
the relations between S.I.A. and the didactics of mathematics consequently
questions the theoretical status that mathematics didacticians can grant to
these models of rules as well as the nature or the status of data from which
the modalities of variables subject to the S.I.A. are constructed.
If one can define mathematical didactics as a plan of scientific study of
the phenomena tied to the transmission of disciplinary knowledge, this means
that the didacticians’ preferred field of observation is that of the mathematics
class: the mathematics class in the sense of the material space, certainly (as
the one can explore the posters, the students’ notebooks. . . ) but also in the
sense of symbolic space (the mathematics class still exists when the teacher
prepares his classes, when the student learns his lessons, at home, at study
hall. . . ). This class thus exists once the interaction between subjects places
them, one as teacher, and the other as student, in relation to an object of
disciplinary knowledge. Very schematically, one of the main objects of study
in of didactics is that of the manifestations of this relationship between these
three interdependent elements, the teacher, the student and the knowledge
of the discipline, of its establishment and its maintenance over time. Two
consequences can be drawn from this modelling. Firstly the observables are
consequently constructed as interactions between master and student, mas-
ter and knowledge, student and knowledge. Secondly, this study requires the
analysis of the regulations that are simultaneously going to be generated by
this system of interactions and assure its functioning. Thus, for example, one
of the most fruitful problems in mathematical didactics is that of the regula-
tions which affect the interactions between the student and the knowledge at
hand if the observables are in this case actions in which the student is engaged
(linguistic or not), the regulations that govern these actions (the engagement
of procedures, some choices made. . . ) or that are produced by these actions
(the abandonment of certain ways of acting, certain controls. . . )
Some of the rules revealed through SIA are thus interpretable in math-
ematical didactics in terms of regulation of interactions. Two positions can
 Didactics of Mathematics and Implicative Statistical Analysis 279

then be adopted: either these rules have the status of hypotheses for the di-
dacticians, it being his responsibility to invalidate them or to confirm them
by other methods of analysis (as for example the undertaking of interviews)
or they have the status of facts of experience (thus allowing the rules to con-
tradict or to not be able to invalidate the analysis a priori ). Some of the
rules revealed through SIA are thus interpretable in mathematical didactics
in terms of regulation of interactions [8, 9].
The asymmetry that these rules present is also to be taken into account
in the interpretation that mathematical didactics can make of them. It poses
the problem of an explication of the asymmetries by the modelling in terms of
a system of interactions. The didactic system that we have summarily evoked
(a triplet of interactions between student, teacher and knowledge) is a system
in which the asymmetries of the characteristics tied by a rule can be explained
in a number of different ways.
Let us begin with the most classic case, in which the established rules
are from data corresponding to observables (the actions or the effective dec-
larations of students or teachers gathered on site). An asymmetry between
observables must correspond to the asymmetry between modalities of vari-
ables linked by a rule which has resulted from S.I.A. “all the subjects having
the characteristic A have the characteristic B”. This asymmetry leads to ques-
tion didactically the fact that very few students have done B without having
done A, have succeeded in B without having succeeded in A, have answered
B without having answered A, that very few teachers have done B without
having done A, etc. These regulations of doing, saying and of their effects
can be the effects of temporality, of differences between the tasks proposed,
of organization of knowledge. . .
Another case can be envisioned, that in which the established rules are
from data corresponding to observables on site, but also from data perpetuated
from these observables. The stability of the latter therefore reveals groups of
“fixed” subjects (the students of a same socio-cultural milieu, “novice” vs. “ex-
perienced” teachers, etc.). S.I.A. can then either provide rules linking actions,
statements, the effects of these actions, of these words and these constituted
groups, or, by the study of the contributions of subjects to rules, establish ten-
dencies shared by subjects from the same group, or on the contrary, equally
characteristic avoidances.
We will present several cases of studies in mathematical didactics exem-
plifying these different uses.

2 Regulations of Situated Actions, Rules Established

from Observable Modalities.
2.1 Asymmetries of Rules Established and Chronology of Tasks
The example that we develop first is that of the study of responses of students
of CM1-CM2 level (9 to 10 years) to an exercise that is composed of two
280 D. Lahanier-Reuter

successive tasks. Firstly students are asked to put into order written decimals
and fractions 1.2; 5.9; 7.5; 4; 9.5; 12; 5.15; 1/2; 2.5 secondly to place them on
a graduated line. The question is: “range par ordre croissant 1,2 – 5,9 – 7,5
– 4 – 9,5 – 12 – 5,15 – 1/2 – 2,5”. In French the word numbers associated
with 5.15 and 5.6 are pronounced ‘five, comma fifteen” and “five, comma six”.
This way of pronouncing the numbers explains a frequent error at this school
level, which consists of placing 5.6 before 5.15, by only comparing the decimal
parts of these numbers. However the numbers have been chosen so that the
reproduction of this classification error in the second part of the task leads
to a contradiction that the students —still normal for students at this school
level— can comprehend. In fact, to put a point corresponding to 5.6 on the
graduated line, then to put one that corresponds to 5.15, in moving back the
first one by a space of “9” (the space between 15 and 6) leads the student to
place 5.15 erroneously on the point that should correspond to 6.5 (5.6 + 0.9).
This placement can seem contradictory with that which corresponds to 6.2
and other points. We will say in this case that the information given to the
student by the erroneous placing of 5.15 is an element of the environment
with which the student interacts.
If consequently we expect certain students to commit errors in the ordering
of written numbers, on the other hand, we wonder about the effects of the
consequences of these errors during the execution of the second task. Two types
of common considerations in mathematical didactics allow us to anticipate
them. Firstly, the information that the erroneous placement of the points
on the line provides is not “naturally” interpreted in terms of a contradiction.
The reading and the comprehension of this information by the student require
that he uses certain knowledge: in fact it concerns considering the placement
of 5.15 and 6.2 as “strange” and sees to consider them as a consequence of the
classification error of 5.6 and 5.15. Previous research done on this error or this
problem in a school situation lead us to differentiate a student’s recognition
of an error from his doing of that error. Or to put it simply, the perception
of a contradiction in his results is often insufficient to lead a student in a
class to invalidate the latter because he still does not feel invested with the
responsibility to resolve the problem raised [1,2,11]. The question of the study
of students’ behaviour is therefore a legitimate question.
The study of the corpus of written productions of the students makes
apparent the diverse strategies used to respond to the two questions of the
exercise. To order the numeric writing, some students used a classification
strategy by ‘types of writing’, in classifying first the written fractionals, then
the written entire numbers, that have no decimal point, then the written
decimals, in separating those that only have one figure after the decimal point
from those that have two. The pupils adopting that classification take into
account only the length of the numbers as they are written out. Others, as we
could have expected, classified the written figures according to their entire part
(visible or calculated in the case of 1/2), then according to their decimal part,
which was also considered as an integer number (5.15 is then placed after
 Didactics of Mathematics and Implicative Statistical Analysis 281

5.6). Finally, certain students ‘neglected’ to use the reference points of the
graduated line and instead used the line as a ‘writing line’ without placing
any points. This study also allows us to decide if, in the end, the student
produces two different orders which are consequently contradictory, or if on
the contrary he produces two coherent orders, even if they are erroneous.
If S.I.A. is applied to data it may result in an association group as in
Graph 1 that allows us to see the following rules:
1. “Adopting, definitively, a classification or writing by types of writing” im-
plies, ‘accepting a lack of accord between the two orders produced” (99%).
(Graph 1, 3 ⇒ 13)
2. “Working, definitively, on the graduated line, as a writing line” implies
obtaining “two coherent orders, even if they are erroneous” (95%) (Graph
1, 7 ⇒ 12).
3. “Producing, finally, an exact classification of written numbers” implies
obtaining “two coherent orders” (95%) (Graph 1, 4 ⇒ 12).
These three rules are interpreted as regulations of student behaviour when
faced with these two tasks. It is possible to read in rule (1) the fact that
certain of these students see —or decide to see— the two tasks as distinct. For
instance, one of the students answers (at first question): “4; 12; 1,2; 2,5; 7,5; 9,5;
5,15; 1/2”. However, he puts marks on the graduated line for “1,2”, next for “2,5”,
next for “4”, etc. We can consider that they do not understand (in the situation
explored) the articulation between the order that the linear arrangement of
the written numbers “shows” and that which the arrangement of points on the
graduated line “shows”. Secondly, rule (2) can be interpreted, in the context
of the situation, as an “avoidance” of the second task. The student copies the
preceding list of writing onto the graduated line, and thus avoids taking into
account the eventual difficulties that he will face in assuring coherence between
the two orders. Finally rule (3), in establishing the asymmetry between the
two modalities1 , leads us to surmise that checking the coherence of the two
orders allows, for certain students, a rectification of the classification of the
written figures.
Thus the didactic organization of these two tasks, and particularly the
conception of an environment by the interpretation of its retro-actions, to
reveal an incoherence of the results, is insufficient: it is necessary in fact, that
the student accepts to link these two tasks in order for him to accept to
interpret the results of one according to the other.

2.2 Asymmetry of Rules and Representations of Subjects.

The use of S.I.A. in didactics of mathematics goes beyond the problematic

that we have just mentioned. Another field of investigation uses S.I.A. as well.

1
That could not particularly be assumed by a test of χ2
282 D. Lahanier-Reuter

We will call usually, the field of reconstruction of observed subjects’ represen-

tations. Indeed, teaching and learning situations can be considered as social
situations defined by the stakes, positions and specific roles of those involved.
The reconstructions the authors of these situations make of these stakes, of
these positions and of these roles have consequences on effective actions. As a
matter of fact, these representations can be considered as knowledge networks.
In that case, S.I.A. can contribute to the recognition of such networks. This
time, implicative rules of the type “if A then B” can be interpreted as follows:
factor A is predominant compared to B in the creation of the representation.
The example that we are presenting here is the study of representations
of some school subjects’ organization for high school students. The notion
of “school subject” is a complex one, even though it can be more or less
naturalized within the school system institution.
What is “French”, for example, or “Physique-Chimie”
(Physics-Chemistry)? Can we describe a school subject by the organization
of knowledge which makes it specific or should we deal with it according to
the teaching and learning techniques that constitute it? If these theoretical
issues are far from being solved, the few studies done with students confirm
the interest researchers have for them.
In fact, it seems that a large number of students have trouble identifying
the different subjects: for example, some of them cannot identify the terms of
a French exercise from the terms of a History exercise. It also seems that the
identification criteria are very often material in primary school pupils (from
7 to 11). Those who best identify the different disciplines say, e.g. they do
so by using material clues, such as a notebook’s colour. However, the point
is that the difficulties have consequences on how well students do in school.
Identifying a discipline’s boundaries and being able to recognize some of its
functioning is a factor of success.
We are presenting here a study about these different issues that deal with
high school students’ representations of scientific disciplines. A questionnaire
was given to four scientific junior classes (Première S in French–), the students
are 17 years old), and scientific senior classes (Terminale S in French –the
students are 18 years old). The questionnaire’s aim was to ask students at the
end of their high school years how they perceive the different scientific fields
they are being taught.
The different questions they were asked are pertinent to the following in-
tensive subjects: mathematics, analysis and statistics identification. Students
of this level recognize analysis and statistics as “parts” of mathematics. How-
ever, what “parts” means is not as clear as it seems to be. Indeed, analysis
and statistics can be considered as separated mathematical fields from an
epistemological point of view: even if some of their objects and methods are
undoubtedly common, knowledge projects, special applications fields, and sym-
bolic representations contribute to specify these scientific areas. But we may
suppose that this approach is not the one the students have. One of our the-
oretical hypotheses is that the partition between analysis and statistics, and
 Didactics of Mathematics and Implicative Statistical Analysis 283

the relationships between mathematics, analysis and statistics, elaborated by

students, are generated by their school practices. Therefore, we intend to focus
especially on differences that can be related to experiences of mathematics,
analysis, and statistics teaching and learning.
The first questions are about the identification of the school level where the
students think the teaching of math (analysis, statistics) started. Students can
usually say in a coherent way when mathematics and statistics started to be
taught (in kindergarten and elementary school for math, in Junior high school
for statistics). Students have much more trouble locating when they were first
taught analysis; they usually avoid the question. Then they are asked how
useful they think the preceding disciplines are. Here again the answers are
pretty clear cut: they see a cognitive use in mathematics, but more rarely any
use in real life (future jobs, etc.). As for statistics as a school subject, it is quite
the opposite, pupils see it as useful outside school, in the real world. The next
questionnaire item asks students to show how mathematics is used in other
disciplines. All students answer that maths are used in physics but very few of
them point to other school subjects in which analysis could be useful. The last
question in this part of the questionnaire, about the usefulness of these three
fields of knowledge, is the following: the student is asked if he remembers or
not his teacher talking about how useful these disciplines can be. The four
last questions concentrate more particularly on school work habits: how they
identify a class (math, analysis, statistics), how they organize themselves (do
they use separated class folders or not), how they identify exercises of the
same subjects, and how much they think they have learnt in these subjects
throughout the school year.
A student’s answer can be considered as an indicator of the way he puts
back together the school subject referred to and the way it is organized for
him, through his memories of it, how useful he thinks it is, and what definition
he gives to it. We therefore consider the answer as a trace of what Yves Reuter
calls “subject awareness” [13].
To understand the graphic (Graph 2) we have only kept the characteristics
of related answers.
The organization into a hierarchy of the different items shows us how the
students deal with the three different disciplines. They have most trouble
identifying analysis whereas they identify mathematics and statistics more
clearly. One of the first results of this study is therefore to show that students
dealing with closely related fields of knowledge, in the same physical space
of the classroom, have some trouble defining the boundaries of the different
fields involved. This result is important for us to know, because in the French
educational system, even in junior high school, students often face classes
like “Histoire-Géographie” (History- Geography), “Physique-Chimie” (Physics-
Chemistry), etc.
Looking at the graph (Graph 2) shows that these identifications are inter-
related.
284 D. Lahanier-Reuter

(1) “Students recognize the characteristics of an analysis lesson” implies

“Students recognize the characteristics of an analysis exercise” (85%) (Graph
2 22 ⇒ 25)
(2) “Students recognize the characteristics to an analysis exercise” implies
“Students recognize a list of the knowledge learnt in analysis class” (90%)
(Graph 2 25 ⇒ 31)
The two combined rules can be interpreted as the traces of a complex
network of knowledge which keeps the representations of a specific subject.
They suggest several levels to these representations of “analysis”: the highest
level would be determined by the ability to identify, in what is taught, the
subject’s characteristics.
Other implicative ways are also to be considered:
(3) “Students recognize the characteristics to an analysis class” implies
“Students recognize the characteristics of a mathematics exercise” (85%)
(Graph 2 22 ⇒ 24).
(4) “Students recognize the characteristics to a mathematics exercise”
implies “Students recognize the characteristics of a statistics lesson” (85%)
(Graph 2 24 ⇒ 23).
The degrees of subject awareness are therefore probably not independent
from one another: the graph nodes in particular (here the identification of an
Analysis class specificities) show the interrelation of these networks.
Lastly, the graphs’ separation also suggests that factors of subject’ iden-
tification are different and not linked. As a matter of fact, the items in the
central graph point back to everything that has to do with reconstruction
within work tasks.
The items on the right end side refer back mainly to what has to do with
reconstruction involving other parties, here the teacher.
(5) “To remember a teacher’s presentation of the usefulness of analysis”
implies “Remembering the presentation of statistics usefulness by a teacher”
(85%) (Graph 2 16 ⇒ 17).
This separation among implicative ways and therefore among networks
making up the representations studied here, is interesting.
What seems to be important here, apart from the fact that what the
teacher says can help define the subjects taught, is on the contrary to find
out that there is no link between subject practices within a classroom and
what the teacher says about them, no link between learning and its social
usefulness.
Subject awareness as it comes up in what students say is not, therefore,
the sum of the different perceived characteristics, but rather the delicate elab-
oration of links between these different characteristics. One of the interesting
aspects of S.I.A. is the way it allows the description of knowledge networks,
since this possibility is directly linked to one of the problematics in didactics,
the study of students’ and teachers’ representations and conceptions.
 Didactics of Mathematics and Implicative Statistical Analysis 285

2.3 Rules Interpreted as Traces of Skills.

One of the last S.I.A. fields of applications in mathematical didactics that

we will present is that of skills reconstruction, stemming from observations of
student behaviours. As we have already seen, rules of the type “if A then B”
showed by S.I.A. must be interpreted.
A and B still refer to observable behaviours. The implicative relationship
can be interpreted, according to the different cases, such as: B is a consequence
of A or A is an explicative factor of B. The reconstruction of students’ ability
to complete a particular task can be read in the implicit rules which govern
students’ behaviours.
The example that we have chosen to explore stems from a study carried
out in the didactics of mathematics, even though it is only a part of a much
wider research which involves different disciplinary didactics [10].
The main problematic in which this study makes sense is in the relation-
ships between teaching and learning, or to be more precise, that of measuring
the effects of a particular pedagogical set of devices on an aspect of mathemat-
ics learning. Numerous prior studies can be quoted on this theme, amongst
which two equally interesting syntheses have been recently and simultaneously
published (on the one hand [12] and on the other hand [3]). Their hypothesis
is that the particularities of didactic management of teaching and learning sit-
uations by the teacher can influence the building of mathematical knowledge
and the appropriation of other knowledge of the discipline by the students con-
cerned. We will extend this hypothesis to attitudes and behaviours students
present when faced with specific tasks.
The study tries to measure the effects of a pedagogical set of devices, which
make it necessary to compare skills developed by students within a particular
study set up and the skills of students who are not in the study setup. We will
get back to our goal later. However, one of the stages of this undertaking, is
first to describe the skills brought forth by the group of students while they
were doing the specific subjects tasks.
The example we will bring out here is, as we have mentioned above, in
this research perspective2 . It is supported by the analysis of skills in geometry
and the language skills of 9 and 10 years old pupils who had to do a writing
geometry task.
The task is the following: “how to draw this figure?” (See Fig. 1).
It was given in seven different classes, by the teachers themselves, without
any outside observer. It thus appeared as a more or less ordinary task within
the class. 165 texts were collected, of which 163 can be taken into account.
The study is therefore about a writing task taken from “an instruction pro-
gram to reproduce a complex figure”. To be carried out, the writer is required
first to identify and to name some of the constructible elements, then to point
out the constructible relationships that exist among the different elements.
2
These results come from an IUFM Research “Effet d’un mode de travail péda-
gogique Freinet en Z.E.P” R/RIU/04/007
286 D. Lahanier-Reuter

Fig. 1. “how to draw this figure?”

We analyze these texts as pupils’ works, that is to say we try to take into ac-
count the context in which they were created and the different pupils’ status.
It is not the same, for instance, for a nine years old or a ten years old pupil to
produce, as a construction program, the following: “draw two perpendicular
lines of which centre is the meeting point of lines and link the intersection of
the two lines and of the circle” or “trace a square, its diagonals and its centre,
trace the circle whose centre is the same as the square’s and which touches
the summits of the square” or “draw a circle and a square inside the circle”. As
a matter of fact, the first text points to objects and geometrical relationships
between the objects that are constructible by pupils at this school level. On
the other hand, the second one needs the drawing of a square, which isn’t as
easy for them. The last one avoids the necessary drawings to be able to fit a
square in a circle. Taking into account the pupils’ school level and considering
the outcomes as productions, we have chosen not to rank them according to
how correct the answers are, but according to the choices pupils made.
We have kept as first indicators of the way pupils write, the chosen el-
ements and their designations, the geometrical relationships mentioned and
their designations. The corresponding indicators are the numbers of geometri-
cal terms used, the chosen elements -circles, lines, summits, . . . -, relationships
such as perpendicular, topological positions, etc. Therefore, the point is to de-
termine the figure analysis pupils have chosen to make through what they say.
The way of analyzing can be quite different from one pupil to the other. Some
of them only dealt with the lines that shape the diagram: the two lines, the
circle and the square, O or the four summits of square A, B, C, D. From a
theoretical point of view, these two ways of looking at it are linked to dif-
ferent analytical skills. As a matter of fact, looking at a geometrical figure
as a punctual structure requires going beyond immediate perception, which
only shows lines entangled. The use of S.I.A. makes it possible to test the
 Didactics of Mathematics and Implicative Statistical Analysis 287

theoretical hypothesis of the different analytical stages in constructing a geo-

metrical diagram.

2.3.1 Split implicative ways: the example of different geometric

skills.
Studying the graph (Graph 3) showing the implicative links between the dif-
ferent items kept makes it possible to dissociate two implicative ways. One of
the networks links the items which take into account points A,B, C, D to those
that determine the diagram’s elements in relation to the others and lastly to
a central item, the one that shows that the pupil not only takes into consider-
ation point O but also gives it at least two different status: for example from
being the centre of the circle, it becomes the lines’ intersection or the square’s
centre. A second network links items which, on the contrary, indicate that the
pupils didn’t take into account the different points (“ONM” and “ABCDNM”)
to those that mark the absence of constraints between the different elements
of the diagram.
The first network groups writing productions where the diagram’s analysis
is rather an analysis in terms of punctual structures. The writing productions
communicate certain constraints, which heavily influence the drawing of the
different elements. The second network gathers texts which are more like de-
scriptions, in which pupils only need to mention visible lines and sometimes
their respective topological positions (inside, on). The first type of output is,
for us, an indication that the skills needed to move on to “a construction pro-
gram task” are met, whereas the second type of output is more the indication
of an interpretation of a task in terms of “description of a regular drawing”.
So, the graph S.I.A. provided, because it can differentiate implicative ways,
allows displaying connected geometric skills separately. We provide here two
texts that are more or less representative of these positions: “Tracez un carré
de 2,9cm. Tracez deux lignes qui se croisent au milieu du carré. Le croisement
des deux droites vient former un centre. Celui-ci (centre) permettra de tracer
un cercle qui touche tous les sommets du carré ” (Valérian, CM2) and “ Il faut
faire comme une croix puis le cercle de 4,2cm et enfin le losange ” (Tiffany,
CM2)3 .

2.3.2 The networks’ nodes: crucial points.

If the dissociation of implicative nodes in a graph is interesting, the identifi-
cation of nodes in the different networks is no less interesting. What we call
nodes here are the items that take part into several links.
In the first network (Graph 4) we notice how important the node is “O
appears and goes through a change of status”. As a matter of fact, pupils can
mention the centre of the circle or the middle of two lines without having to
get into a perspective of linking both lines (or even saying it).

3
Original spelling has been changed.
288 D. Lahanier-Reuter

Therefore it is indeed the change of status (in turn, centre of the circle and
intersection of lines or centre of the circle and intersection of the diagonals)
which makes up the decisive criterion for classifying the pupils’ production
and its grading.
Trusting the reader with this change and finding the discursive ways to do
so might be a crucial stage. It involves indeed being able to find the means
to bring back, to recall an element already present in the text (therefore a
handling or the implementation of anaphora), it also implies being able to get
away from visual evidence of the figure, in short being able to go beyond an
immediate visual contact to a written account of an invisible change. Point O
doesn’t move but its function changes.
So S.I.A. allows us proving the deciding role played by some analysis cri-
teria and the necessity of their presence in skills evaluation.

2.3.3 Univocal implicative links: the case of some linguistic

characteristics.
The complexity of some of the graphs studied above shouldn’t mask the fact
that in some cases they are in fact extremely simple. Their “simplicity” is
nonetheless a source of information that shouldn’t be overlooked. Here we will
focus on the study of links between the items corresponding to the linguistic
characteristics of the texts produced by pupils (see Graph 3).
These characteristics have to do with the length of text produced, the
different modes used (infinitive, imperative, indicative), subjects (“I”, “we”,
“you”). We have also kept the signs of planning: as a matter of fact the task
that is proposed can be interpreted as one of the writing of a series of actions
aimed at putting the figure back together. Some pupils write in an orderly
list of actions. Others note the temporality by using adverbs (now, then, . . . ).
Others conclude their texts by an indication of the type “here it is, the figure
is done” or simply by using the word “end”. Some disorders can be caused by
planning operations. Thus, some pupils refer to elements that have not yet
been introduced in their text; others add constraints that “they had forgotten”.
Lastly, one of the last characteristics of the produced writings is that of the
inadequate use of definite articles — “the. . . ” and indefinite — “a, some. . . ”
As a matter of fact, the presented elements can be undetermined by what
precedes them or on the contrary entirely determined. For instance, if the
pupil has said how to build the four points (and give name to them ABCD)
on two perpendicular lines which intersection is 0, the circle he is then going
to talk about (0 being its centre and going through the 4 points) is entirely
determined. Thus these determinations are not of a linguistic order: it isn’t
because one element has already been quoted in the text that it is therefore
determined, but because the geometrical constraints define it in a unique
way. There is no doubt then, that the tension between the two orders of
determination explains the numbers of disorders in the use of articles.
Even though there are quite many linguistic characteristics, on the other
hand, the graph isn’t really complex. Three main rules stand out:
 Didactics of Mathematics and Implicative Statistical Analysis 289

• (1) “Writing “I” implies “using the indicative mode” (99%) (Graph 3
12 ⇒ 11).
• (2) “Using the infinitive mode” implies “Building a generic subject “one”
“(99%) (Graph 3 9 ⇒ 13).
• (3) “Using the imperative mode” implies “building a subject “you” ” (99%)
(Graph 3 10 ⇒ 14).
The links that come up are expected for the most part since the very use,
even partial, of the imperative and infinitive modes is indeed linked to the
pronouns which signal what the reader puts back together: a “peer” for the
imperative mode, signalled by the pronoun “tu” (informal you) and “vous”
(formal you), “a generic reader” for the infinitive mode, an “evaluative reader”
signalled by “I” and the indicative mode. However, the presence of univo-
cal relationships between the chosen pronouns and the modes used, gives an
unexpected rigidity since it is possible, in the indicative mode to use “on”
(people)” and “tu” (informal you)”. Unlike classical tests suggesting symmet-
rical links between studied variables, the S.I.A. allows questioning of these
strong constraints, stemming from the fact that the writing is produced in a
school situation.
This makes it possible to think that the rules students use define them
as actualisations of discursive genres. Thus, we suppose that building the
reader as a “peer” is characteristic of a school writing genre in math class, and
therefore can be perceived as legitimate by pupils for several reasons. It could
be that the pedagogical and didactic devices make such positions possible,
because help and cooperation are principles put into practice in the language
used in the different subjects taught. It could also be that the ways exercises
are written in school books define such a reader. Building a “generic reader”
is also an identifiable characteristic. However, this characteristic is not as
frequent in geometric construction exercises in elementary school schoolbooks.
On the other hand, it is frequent in “description of recipes” schoolbooks.
At this level, this genre, like those of “construction programs” or “users’
manuals”, take up a rather important part in school activities, and can be
found in places other than schools. Lastly, the discursive ways of using “I”
with which a pupil shows the reader what he can do or manages to do is a
characteristic of school evaluation situations.
Therefore, S.I.A. interpretation of results appears to take into account
univocal links. Unlike in traditional analysis, links can be interpreted as con-
straining rules governing the actions observed.
290 D. Lahanier-Reuter

3 Regulations Relative to Groups of Subjects, Rules

Established in Observable Modalities
and in Contributory Variables Modalities.
An important problematic in mathematics didactics is, as we have shown,
that of the interpretation of regulations in pupils’ actions seen in a group’s
characteristics. It is a matter of trying to know if we can give the status of
“results of teaching devices used” to certain behaviours and to certain skills.
These issues are what trigger experiments as well as comparisons of ordinary
practical school experiences. It is about sorting out what is specific to a group
of pupils, whichever methodology is used: in the case of experiments, we are
trying to compare the performance or competence of pilot groups and those
of experimental groups, in the case of observations, the groups constituted are
all of the classes.
The study mentioned above addresses this question, since it questions
some teachers’ demands to set up specific teaching devices in their classes.
The study attempts to describe the effects of such a way of functioning, from
the point of the pupils’ performance. Its goal is to show the results strengthen-
ing or invalidating the hypothesis according to which, particular effects, read
in pupils’ behaviours, can be linked to specific devices used in classes. It is
this question that brings us to compare the readings done by the researcher
who studied the pupils’ group activities, according to whether the students re-
ceived one kind of teaching or another. Remember that there are seven classes
studied. They are elementary classes respectively 9 year olds and 10 year olds,
all located in the suburbs of Lille.
Taking up again the different characteristics studied, relative to geometrics
and linguistics skills, we are now studying links connecting the different items,
showing to what class they belong.
This graph (see Graph5) has been completed keeping only the paths lead-
ing or getting to one of the explored “classes”: the different CM1 (9 year olds)
or CM2 (10 year olds) classes. We are thus trying to bring forth characteristics
in groups of pupils.

3.1 “Isolated” Groups

Methodologically, we can first look at the isolated groups. In the case of the
example we are developing, one of the CM2 class (“CM2 Wa”) is isolated. As
it would be the case in “classical” analysis, the absence of implicative links
marking this group of pupils is interpreted as a sign of diversity in this group
of pupils’ written productions (according to the chosen criteria).

3.2 Characteristic Abilities

S.I.A. makes it possible to perceive the cases where characteristics are those of
a part of the group of pupils studied, compared to a group where all the pupils
 Didactics of Mathematics and Implicative Statistical Analysis 291

of the group were studied. They “almost” all share the same characteristics,
contrary to the modes of analysis which produce symmetrical links. We will
start by presenting a case where a characteristic comes up as specific to one
of the groups.
The example is that of the characteristic “writing a text using “I” which is
the characteristic of one of the classes coded CM1 Wb.
- (1) “Writing a text using “I” ” implies “belonging to class CM1 Wb” (99%)
(Graph 5 12 ⇒ 2).
Only the pupils of this class chose a particular writing behaviour, which
indicates that pupils read into a suggested situation an evaluation situation:
the expected reader is the teacher and pupils show what they can do. But what
seems even more important is the fact that the implicative link of maximal
intensity means that no pupil “or rather almost none” in other classes have
reacted in that way. To explain this specificity, we support the hypothesis
that the way the task was done has been singularized in class: maybe only
this CM1 teacher has presented the exercise as an evaluation or at least as an
exercise that he would check on.

3.3 Groups Characterized by Capacities

On the other hand, other implicative links show that almost all pupils of a
same class share some skills and also make the same mistakes, etc. Let’s keep
in mind some of them:
• (1) “Belonging to CM2 Hb1” implies “using “tu” ” (90%) (7 ⇒ 14).
• (2) “Belonging to CM2 Wa” implies “Mentioning point 0 and changing its
status” (95%) (5 ⇒ 29).
This time, these characteristics are met in almost all the pupils of a same
class.
We think that they are the results — sometimes indirect — of the didactic
or pedagogical approaches used in class. What is left is to interpret these
different rules. If almost all the pupils of the first particularized class address
a reader who acts like “a peer” in their writings, it is no doubt because help and
cooperation are legitimized and encouraged in these classes or because these
forms of communication are used. If almost all the pupils of the second class
show good geometric skills, it is no doubt thanks to the teaching techniques
used.
But this geometry skill cannot be understood without the linguistic skills
which make it possible to communicate it. Using “tu” (informal “you”) is not
always the required school form. Now, these two classes have pupils from
different social backgrounds: in the second class the pupils come from more
privileged families than the other studied classes. Relationships between so-
cial classes and linguistic strategies are certainly complex and not mechanical.
However, the results we are getting are coherent with those of other studies
292 D. Lahanier-Reuter

on these relations. As a consequence, we cannot neglect the explicative factor

of cohesion between observed behaviours. We are bumping here into a recur-
rent problem: the principles of subjects’ categorisation are obviously never
unique or uniform. The gatherings of pupils in this particular case cover both
institutional groups (school classes) and “social” groups (coming from more or
less privileged families). This remark makes it very clear that evaluating the
effects of teaching methods cannot take the shape of simple cause to affect
relationships.
Another way for bringing up class characteristics is provided by S.I.A. We
may have access to contributions of every subject or of additional variables to
implicative links [7]. These methods allow us to count how many pupils of each
class contribute to these significant links, or to exhibit which group of subjects
has more weight in these established links. So doing, we can confirm some of
the previous exposed results and, on the other hand, bring to the fore some new
ones. An example of confirmation of previous results is given by looking at the
subjects that contribute to the link “using “I” ” and “using indicative mode”.
Not surprisingly, all these subjects belong to the CM1 Wb class. Considering
the additional variables “belonging to CM1 Wa”, “belonging to CM1 Wb”,
“belonging to CM1 Hb1”, . . . , S.I.A reveals that “belonging to CM1 Wb” is the
most contributive variable to the link “using ‘I” and “using indicative mode”.
Analyzing contributions may also produce some new results. For instance, the
implicative link between “square located” and “ABCD not mentioned” (see
Graph 4) may be interpreted as the following: some of these pupils regard
the given figure as a network of lines and not as a structured set of points.
Analyzing contributions of additional variables mentioned above affords no
determining difference between them. Nevertheless, it appears that 66% of
the optimal group are CM1 pupils, considering the subjects’ contributions to
this link. So, a more adequate additional variable to this part of the study is:
“belonging to CM1 (or not)”, which, actually contributes mostly to this link.
This being said, it is in any case obvious that using S.I.A., even if it doesn’t
help to find cause and effect relationships, makes it possible to accumulate,
little by little, shared behaviours, common capacities, specific mistakes, . . .
linked to groups of subjects. What is left to find, then, is global coherence to
the groups’ characteristics.

4 Conclusion
Using a central problematic in mathematical didactics, we have been able
to show the efficiency of dealing with S.I.A. techniques. The input of this
method to analyze data cannot be disregarded for several reasons. The almost-
rules established by S.I.A. from observables can easily lend themselves to
interpretation in terms of action regulation. The implicative paths can be
read in terms of networks. Lastly, the asymmetry of links seems essential
in exposing any explicative hypothesis of certain phenomenon pertinent to
 Didactics of Mathematics and Implicative Statistical Analysis 293

teaching and learning. We have also been able to sketch, through the detailed
relationships of research examples, particular methodological behaviours: the
attention to give to the interpretation of the internal cohesion of implicative
graphs, but also to the separation of these graphs, as well as to the graphs’
nodes and the univocal links. These are paths of thinking to be pursued.

References
1. G. Brousseau. Le contrat didactique : le milieu. Recherches en didactique des
mathématiques. Volume 3, pages 309–336, La Pensée Sauvage, Grenoble, 1990.
2. G. Brousseau. Théorie des situations didactiques. La Pensée Sauvage, Grenoble,
1998.
3. M. Bru, M. Altet, C. Blanchard-Laville A la recherche des processus caractéris-
tiques des pratiques enseignantes dans leurs rapports aux apprentissages. Revue
Française de pédagogie. Volume 148. INRP, Paris, 2005.
4. R. Gras. L’analyse des données : une méthodologie de traitement de questions
de didactique. Recherches en didactique des mathématiques. Volume 12:1, pages
59–72, La Pensée Sauvage, Grenoble, 1992.
5. R. Gras, A. Totohasina, S. Almouloud, H. Ratsimba-Rajohn, M. Bailleul. La
méthode d’analyse implicative en didactique. Applications. In : M. Artigue,
R. Gras, C. Laborde, P. Tavignot (eds.): Vingt ans de didactique des mathéma-
tiques en France. Pages 349–363, La Pensée Sauvage, Grenoble, 1994.
6. R. Gras. L’implication statistique, Nouvelle méthode exploratoire de données.
La Pensée Sauvage, Grenoble, 1996.
7. R. Gras, J. David, J.C. Régnier, F. Guillet. Typicalité et contribution des sujets
et des variables supplémentaires en Analyse Statistique Implicative. Extraction
et Gestion des Connaissances (EGC’06). Volume 2, pages 359–370, Cépaduès
Editions, 2006.
8. D. Lahanier-Reuter. Conceptions du hasard et enseignement des probabilités et
statistiques. P.UF., Paris, 1999.
9. D. Lahanier-Reuter. Exemple d’une nouvelle méthode d’analyse de données :
l’analyse implicative. Carrefours de l’éducation. Volume 9, pages 96–109, CRDP
Amiens, 2000.
10. D. Lahanier-Reuter. Enseignement et apprentissage mathématiques dans une
école Freinet. Revue Française de Pédagogie, Volume 153, pages 55–65, INRP,
Paris, 2005.
11. C. Margolinas. De l’importance du vrai et du faux. La Pensée Sauvage, Grenoble,
1993.
12. A. Mercier, C. Buty C. Evaluer et comprendre les effets de l’apprentissage de
l’enseignement sur les apprentissages des élèves : problématiques et méthodes
en didactique des mathématiques et des sciences. Revue Française de Pédagogie.
Volume 48, pages 47–59, INRP, Paris, 2004.
13. Y. Reuter. Les représentations de la discipline ou la conscience disciplinaire. La
Lettre de la DFLM. Volume 32, pages 18–22, 2003.
294 D. Lahanier-Reuter

Appendix
a) S.I.A Graph 1.

Chronology of tasks, two types of number classifications.

Fig. 2. 1 Writing classification: No answer; 2 Writing classification based on nu-

merical relationship with errors; 3 Writing classification based on length; 4 Writing
classification exact; 5 Writing classification incomprehensible; 6 Writing classifica-
tion with inversion 5,15 and 5,9; 7 Linear classification based on ‘line’; 8 Linear
classification No answer; 9 Linear classification based on points; 11 Linear classifi-
cation with inversion 5,15 and 5,9; 12 Adequation between the two classification; 13
Non adequation; 15: Linear classification with 5,15 pointed on 6,5.
 Didactics of Mathematics and Implicative Statistical Analysis 295

b) S.I.A Graph 2.

Questionnaire and students representations of disciplines.

Fig. 3. 6 Analysis first teaching identified; 9 Mathematics cognitive use identified;

10 Mathematics social use identified; 11 Analysis cognitive use identified; 15 Re-
membering teacher speech about Mathematicsmathematics usefulness; 16 Remem-
bering teacher speech about Analysis usefulness; 17 Remembering teacher speech
about Statistics usefulness; 19 Identification of Analysis use in other disciplines;
20 Identification of Statistics use in other disciplines; 21 Identification of Math-
ematics lessonmathematics class characteristics; 22 Identification of Analysis les-
son classcharacteristics; 23 Identification of Statistics lessonclass characteristics; 24
Identification of Mathematicsmathematics exercises characteristics; 25 Identification
of Analysis exercises characteristics; 28 Using a special notebook for Analysis; 31
Analysis knowledge identified; 32 Statistics knowledge identified.
296 D. Lahanier-Reuter

c) S.I.A Graph 3.

Writing in geometry (all items included).

Fig. 4. 1 CM1 Wa; 2 CM1 Wb; 3 CM1 Hb; 5 CM2 Wa; 6 CM2 Hb1; 77 CM2 Hb2;
9 Using infinitive mode; 10 Using imperative mode; 11 Using indicative mode; 12
Using ‘Je’ (I); 13 Using ‘On’ (one); 14 Using ‘Tu or Vous’ (informal you or formal
you); 15 Planning marks; 16 End marked; 17 Error on ‘le’ (the) or on ‘un’ (a); 18
Circle determined; 19 Circle independent; 20 Circle located; 21 Square independent;
22 Square located; 23 Square determined; 24 Lines independent; 25 Lines located; 26
Lines determined; 27 O No mention; 28 O mentioned; 29 O mentioned, two status;
30 ABCD No mention; 31 ABCD mentioned; 32 ABCD mentioned, two status; 33
ABCD constructed.
 Didactics of Mathematics and Implicative Statistical Analysis 297

d) S.I.A Graph 4.

Writing in geometry, geometrical abilities.

Fig. 5. 18 Circle determined; 19 Circle independent; 20 Circle located; 21 Square

independent; 22 Square located; 23 Square determined; 24 Lines independent; 25
Lines located; 26 Lines determined; 27 O No mention; 29 O mentioned, two status;
30 ABCD No mention; 31 ABCD mentioned; 32 ABCD mentioned, two status; 33
ABCD constructed
298 D. Lahanier-Reuter

e) S.I.A Graph 5.

Writing in geometry, groups characteristics.

Fig. 6. 1 CM1 Wa; 2 CM1 Wb; 3 CM1 Hb; 5 CM2 Wa; 6 CM2 Hb1; 77 CM2
Hb2; 10 Using imperative mode; 11 Using indicative mode; 12 Using ‘Je’ (I); 13
Using ‘On’ (one); 14 Using ‘Tu or Vous’ (informal you or formal you); 15 Planning
marks; 19 Circle independent; 21 Square independent; 24 Lines independent; 27 O
No mention; 29 O mentioned, two status; 30 ABCD No mention.
Using the Statistical Implicative Analysis
for Elaborating Behavioral Referentials

Stéphane Daviet1,2 , Fabrice Guillet2 , Henri Briand2 , Serge Baquedano1 ,

Vincent Philippé1 , and Régis Gras2
1
PerformanSe SAS, Atlanpole La Fleuriaye, 44470 Carquefou,
{stephane.daviet, serge.baquedano, vincent.philippe}@performanse.fr
http://www.performanse.com
2
LINA-École Polytechnique de l’Université de Nantes,
La Chantrerie – BP 50609 – 44306 Nantes CEDEX 3
{stephane.daviet, fabrice.guillet, henri.briand,
regis.gras}@univ-nantes.fr

Summary. Various informatic assessment tools have been created to help human
resources managers in evaluating the behavioral profile of a person. The psycholog-
ical basis of those tools have all been validated, but very few of them have follow
a deep statistical analysis. The PerformanSe Echo assessment tool is one of them.
It gives the behavioral profile of a person along 10 bipolar dimensions. It has been
validated on a population of 4538 subjects in 2004. We are now interested in building
a set of psychological indicators based on Echo on a population of 613 experienced
executives who are 45 years old and more, and seeking a job. Our goal is twofold:
first to confirm the previous validation study, then to build a relevant behavioral
referential on this population. The final goal is to have relevant indicators helping
to understand the link between some behavioral characteristics and current profiles
that can be categorized in the population. In the end, it may provide the founda-
tion for a decision support tool intended for consultants specialized in coaching and
outplacement.

Key words: Statistical Implicative Analysis, Assessment tool, Behavioral referen-

tials, Decision support system, Validation study

1 Introduction
Human resources managers have been early users of computer tools. The need
for evaluating the behavioral profile of a person in human resources has led to
the creation of personality assessment tools. Initially, the first tools were paper
based ones. Then, the first developed computer ones were expert systems (e.g.
Human Edge [12]). Then more complex decision support tools were created:
MBTI (Myers-Briggs Type Indicator) [5, 14], PerformanSe Echo. Meanwhile,

S. Daviet et al.: Using the Statistical Implicative Analysis for Elaborating Behavioral
Referentials, Studies in Computational Intelligence (SCI) 127, 299–319 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
300 S. Daviet et al.

great strides were accomplished in the field of knowledge discovery in data

(KDD) [6], enabling the study of the huge bulk of data collected by those
assessment tools.
Today, validation has become a crucial stake for those tools. Till now, they
were only validated by the relevance of their results and how they matched
the a priori knowledge of psychologist experts on specific assessed stereo-
typed people. Probing them with less subjective methods is yet crucial for
both the assessor firms and the assessed people. Very few studies were driven
to confront personality assessment tools with the reality they are meant to
model. The results of these studies on the MBTI tool was conflicting: [2, 19]
versus [5, 11, 20]. Very few studies have been driven on the Big Five theory,
and on the Echo PerformanSe tool (that we are interested in). Yet, the Per-
formanSe tool is widely used and we have at our disposal of a huge population
to manage some relevant statistical studies to validate it. Mining the sample
of assessed population for association rule discovery [1] seems to be a right
way to achieve this validation.
A previous validation study has been realized over a scope of 4538 evalua-
tions [16,17]. In our case, we are interested in the data collected by the APEC3 .
More specifically, we have carried out a study on a lower sample of population:
the executives who are 45 years old and more and seeking a job, extracted
from a higher population of 2788 people. We have used the CHIC [4] software
and the Statistical Implicative Analysis to conduct our study. We have used
classical statistical measures like mean and standard deviation, and the tools
of SIA [9]: similarity trees [13], implicative trees and cohesitive graphs [10].
Study targets four objectives. First, combined with the previous validation
study of the Echo tool, a statistical survey of this population could confirm
the last collected results. Then, we want to draw a deep statistical analysis
of this specific population to build a referential on which we can establish a
decision support system. The final goal is to have relevant indicators helping
to understand the link between some behavioral characteristics and current
profiles that can be categorized in the population. In the end, this may provide
the foundation for a decision support tool intended for consultants specialized
in coaching and outplacement.
First, we describe the data we have studied. Then, we present our method-
ology to qualify behavioral indicators. We use CHIC to highlight some com-
binations of characteristic behavioral dimensions. Then, we focus on some of
these combinations to bring out the indicators and to associate an appropri-
ate meaning based on expert evaluation to each of them. Second, we present
our results in terms of relevant behavioral indicators. Third, we discuss the
possibility of completing this approach with a temporal analysis. Finally, we
open some possible paths to improve this work.

3
APEC stands for Agence pour l’Emploi des Cadres, in English: Job Center for
Executives
 Elaborating Behavioral Referentials with SIA 301

2 Applicative Context

Various personality assessment tools are widely used in human resources man-
agement for profiling people in job oriented decision support. A personality
assessment tool intends to draw up the behavioral profile of a person from
the results of a questionnaire. The goals of those types of tools are multi-
ples: support for recruitement, support for vocational guidance, behavioral
checkup accompanying a competence checkup. Those tools are not intended
to be used in a discriminative way to select among applicants for a job, but
more as a basis to help a human ressource manager for instance when receiv-
ing people for an interview. We have two types of questionnaires. The first
one is composed of open questions where the person is free to enlarge. The
answers are examined by a psychological expert to draw up the behavioral
profile. It is highly questionnable due to the subjectivity and variability of the
interpretation of the expert from one person to another. There are very few
questionnaires of this type. To name but one, Phrases [18] has the attendee
complete 50 phrases in 30 minutes, under the scrutiny of the examiner. Both
the answers and the behavior of the person during the test are evaluated. This
type of questionnaire is poorly studied due to the difficulty to build statistical
analysis on open questions.
The second type of questionnaires is composed of closed questions and is
the most widespread. Those questionnaires could be handwritten or comput-
erized. It generally consists of a set of questions (also named items) with 2 or
more answers. A set of rules, like those we can encounter in expert systems,
have beforehand been established and give a behavioral profile along a pre-
determined number of personality traits (also named dimensions). There is a
great number of those types of tools. Here we show a set of the computerized
ones:
• Sosie (from ECPA): 20 personality traits evaluated through 98 groups of
4 assertions,
• PAPI (PA Preference Inventory from Cubiks):
– the classic test: a choice between 90 pairs of sentences,
– the normative test: 126 assertions with a choice from “totally disagree”
to “totally agree”,
• MBTI (from Myers and Brigg): 126 questions with a choice between two
answers and a profile chosen among 16 predefined ones,
• PerformanSe Echo: 70 questions with two answers and a profile on 10
bipolar dimensions determined through a set of rules,
• Assess First: 90 questions with two choices drawing a profile over 20 be-
havioral dimensions and 5 families.
Among all these tools, few of them have undergone a real statistical val-
idation study. Indeed, most of those products are based on well-grounded
psychological basis: Jungian theory [5, 15], Big Five model [8, 21], or study
of motivations [7]. But statistical studies are a necessary counterpart of the
302 S. Daviet et al.

psychological validation. Studying the distribution of the population over the

behavioral dimensions could, for instance, be an interesting type of valida-
tion. Several studies have been driven on the MBTI tool, but the results are
conflicting. Some studies [2, 3, 19] have shown that the MBTI is a valid and
reliable instrument, others have demonstrated several drawbacks [5, 11, 20].
The PerformanSe Echo tool is the one we have studied. Previous validation
studies have already been realized on this tool every 5 or 6 years since 1985.
The last one [17] dates from 2004 and it consists in analysing and fine tuning
the distribution of 4538 assessments over the 10 behavioral dimensions of the
PerformanSe model. It results from the collaboration of the KOD (KnOwl-
edge and Decision) laboratory of Polytech’Nantes, the DPL (Development
Psychology Laboratory) laboratory of Rennes 2, the LRI CNRS laboratory
from Orsay and the PerformanSe company. The first goal of this study was
to get a global overview of the repartition of the population among the 10
dimensions. Indeed, with the time passing and the evolution of certain factors
(environment, vocabulary, conceptual references, etc.), it becomes necessary
to control that the questionnaire and its results are always up to date and rel-
evant, and to make it evolve if needed. This is the second goal of this study: to
recalibrate if needed the tool over this reference sample of 4538 assessments.
We will explain in the next chapter concerning the data studied how the tool
is calibrated.
In this paper, we have driven a second and more specific study on a sub-
sample of the population. We focus on a cross section of the 4538 evaluations
which concerns the executives who are 45 years old and more and seeking a
job. The data have been collected by the APEC and anonymized.

3 The PerformanSe Echo Tool

The personality assessment tool, Echo, developed by PerformanSe is a ques-

tionnaire with 70 items. Each item is a question with two possible answers,
but it is not a Yes/No questionnaire: it is called an ipsative questionnaire. For
instance, Fig. 1 shows a question of Echo and its two answers.
Once all the items are answered, the tool draws up the behavioral profile
of the person. This profile is described along 10 bipolar dimensions which are
detailed in Tab. 3. Each pole of a dimension is called a trait and is valued
on a scale from 0 to 35. Each answer ascribes a set of points to one or more
traits. From the scores of two opposite traits, one calculates the score of the
corresponding bipolar dimension. Each of these dimensions, initially gradu-
ated from 0 to 100, has been then discretized in 3 zones: low values under
40 (marked -), medium values between 40 and 60 (marked 0) and high values
above 60 (marked +). For instance, for the extroversion, we distinguish EXT-,
EXT0 and EXT+. Fig. 2 gives an example of a behavioral bipolar profile.
Each dimension matches a personality trait which has no intrinsic real-
ity. This is the interaction between several traits that trigger an observable
 Elaborating Behavioral Referentials with SIA 303

Introversion (INT) Extroversion (EXT)

Express: reserve, modesty, discretion, risk of Express: expansion of self, desire to be no-
looking cold, difficulty to communicate, abil- ticed, ease of expression, risk of attention
ity to concentrate scattering, tendency to be invasive
Relaxation (REL) Anxiety (ANX)
Express: maintenance of a state of relaxation, Express: pressure, worry, emotive power,
cold-bloodedness maintenance of a waking state, concern, state
of tension
Questioning (QUE) Assertion (ASN)
Express: concern for improving, level-headed Express: self-confidence, innermost convic-
opinions tion, firm opinions
Determination (DET) Receptiveness (REC)
Express: distance regarding others, emotions, Express: opening towards others, taste for lis-
passive resistance tening to others, understanding (empathy)
Improvisation (IMP) Rigor (RIG)
Express: taste for the unforeseen and adapta- Express: structure of work and environment,
tion, spontaneous reaction to events, impul- sense of method and planning, sense of hier-
siveness archy
Intellectual conformism (INC) Intellectual dynamism (InD)
Express: reference to well-tried solutions, an- Express: creativeness, social relationships,
alytical approach, sense of precision,difficulty intellectual curiosity (new ideas), overall un-
to take a global view of situations, expert derstanding of situations, quick-wittedness,
knowledge risk to neglect details, overall views
Conciliation (CCL) Combativeness (COM)
Express: patience, search for serene relation- Express: reactive behaviour,search for stakes,
ships, spirit of consensus, ability to act as an sense of competition, offensiveness, impa-
arbiter tience
Motivation for faciliation (FAC) Motivation for achievement (ACH)
Express: immediate pleasure Express: persevering and succeeding
Main fear: having too much work Main fear: being obliged to give up
Stimulus: easiness, short missions Stimulus: difficult projects
Satisfaction: achieving easy success Satisfaction: making efforts
Money rewards for: jumping at the opportu- Money rewards for: merit
nity
Relationship to time: short term projects Relationship to time: long term projects,
feels guilty when losing time
Voc.: to save time, to use short cuts, to give Voc.: to build, to persevere, to deserve,
greater importance to the present. . . tenacity. . .
Motivation for independence (IND) Motivation for belonging (BEL)
Main fear: being overwhelmed by the group Express: influence
Stimulus: personal freedom Main fear: being expelled
Satisfaction: having one’s own territory Stimulus: the community
Money rewards for: individual results Satisfaction: living in good relationships
with people
Relationship to time: protects one’s personal Money rewards for: common results
time
Voc.: to take the consequences of one’s own Relationship to time: dedicate time for the
choices. . . group
Voc.: consensus, solidarity. . .
Motivation for protection (PRO) Motivation for power (POW)
Main fear: not having any guarantee Express: risk-taking
Stimulus: maintenance of what has been ac- Main fear: having no influence
quired
Satisfaction: peacefulness of mind Stimulus: challenge; deciding and leading
Money rewards for: an acquired right Satisfaction: initiating events
Relationship to time: is provident Money rewards for: risk and responsibility-
taking
Voc.: to stay in a known environment, to Relationship to time: wants to leave a mark
avoid surprises, to perpetuate organisation
Voc.: to be in a dominant position, to be am-
bitious. . .
Table 1. The 10 behavioral dimensions
304 S. Daviet et al.

Fig. 1. Screenshot of an ipsative question

Fig. 2. Screenshot of the behavioral bipolar profile along the 10 dimensions

behavior. Such a set of dimensions is called a factor: a meta-concept linked

to a general theory of the personality widely validated and recognized (for
instance, Agreeableness in the Big-Five theory). The tool consists of 27000
rules that draw up a human readable text report from a text base containing
2500 pages of text. This text report considers the combination between each
dimension to give the real behavioral explanation of the profile. Fig. 3 shows
a schema explaining how the system works.
This model is based on the so-called “Big Five” model which describes per-
sonality according to five dimensions: Extroversion, Conscientiousness, Agree-
ableness, Openness and Emotional stability. It is the result of more than forty
years of work led by dozens of researchers: Cattell, Fiske, Eysenck, Gulford,
Tupes, Christal and Norman, and more recently, Smith, Borgotta, Goldberg,
Mc Crae and Costa. This well-tried model has been enhanced by:
 Elaborating Behavioral Referentials with SIA 305

• the study of motivations and what leads the individual to act,

• the systemic and behavioral approach that takes an individual and its
interactions with the environment as a whole.

Fig. 3. Explanation of the system of rules

4 Problematic and Goals

This study has been governed by a need of the APEC to get relevant behav-
ioral indicators for supporting their everyday task: to help people finding the
right job. The APEC is a national organization that provides job guidance
to people. It is somewhat comparable to the ANPE4 , but especially intended
for executives. They provide an assistance for all that is related to job ori-
entation, training, reemployment, skill validation and skill assessment. Their
major difficulty is to get the good clues to determine at best the function
that matches a given person. The PerformanSe tools have been purpose-built
for the employment field. The assessment provides a set of recommandations
with support/vigilance points. But, those conclusions are quite general and
the need of the APEC is more specific to each position. The stake is twofold:
first, to determine the main characteristics that promise the better chance
of success for reemployment, then the personal profile that better matches a

4
ANPE stands for Agence Nationale Pour l’Emploi, in English: National Job
Center
306 S. Daviet et al.

given job. Finding those characteristics boils down to building a job referential
which is our main goal.
To meet those needs, we have organized our study in two steps: the first
step consists in bringing out the specificities of this population of executives
in respect to the global population. To succeed in this task, both classical
statistical tools and more advanced SIA tools are used. In the end, we want
to determine both the dimensions and the factors (combination of multiple
dimensions) that characterize this population, and thanks to a psychological
expert give a meaning to the discovered factors. As previously said, it is not
the dimensions themselves but the factors that can be interpreted. The SIA is
a good way to obtain those combinations of dimensions, notably via similarity
trees and cohesive graphs.
This first study would bring some indicators that may differentiate the
studied population from the global population, but also discriminate some
subgroups in this studied population. If global indicators may characterize
the main part of the sample, there might be some subpopulation that is not
fully characterized by those indicators and could be interesting to study. To
summarize, the first step may bring the discovery of subgroups that we will
analyze in a second more local step. We will use the same tools of usual
statistics and SIA to drive this study.

5 Data

5.1 The Reference Population

The reference population is the one that has been used for validating the tool.
The data has been collected through a partnership between PerformanSe and
a large sample of clients, who have communicated their assessments. It is
composed of 4538 people with wide-ranging backgrounds:
• companies, national and international groups and SMEs in all sectors of
the economy,
• consultancy firms,
• business schools and engineering schools,
• public organizations for professional mobility and orientation, governed by
the Ministry of Employment or the Ministry of Education.
People within this sample could be of any age, employed or not, from mis-
cellaneous socio-cultural origins. Each person of the sample is described with
the 20 traits of the Echo questionnaire. For the computation, it is the val-
ues from 0 to 35 for each trait that have been used, not those of the bipolar
dimensions. On each of the 20 traits of the Echo model, the average score
values of this sample teeters from 17.12 to 17.98 on a scale from 0 to 35. The
standard deviation values are spread between 5.83 and 7.42. The population
follows a normal distribution (centered Gaussian) over the 20 traits. People
 Elaborating Behavioral Referentials with SIA 307

are divided out: 25% in low values, 50% in medium values and 25% in high
values. Tab. 5.1 shows the results of this study over the 20 traits. On every
trait, the population follows a centered Gaussian distribution. It is important
to specify that this distribution has been obtained directly from the gross
results without any curve fitting. That shows the relevance and accuracy of
this personality assessment tool. This study has also shown that the tool did
not need to be recalibrated.

Traits Mean Standard deviation Minimum Maximum

EXT 17.3642 7.0355 0.0000 35.0000
INT 17.4916 6.8063 0.0000 35.0000
COM 17.8043 6.8395 0.0000 35.0000
CCL 17.1670 6.0825 0.0000 35.0000
ANX 17.4746 7.2266 0.0000 35.0000
REL 17.2259 5.9921 0.0000 35.0000
ACH 17.9843 7.3375 0.0000 35.0000
FAC 17.8944 6.7562 0.0000 35.0000
InD 17.9006 6.1809 0.0000 35.0000
InC 17.3321 7.0624 0.0000 35.0000
RIG 17.3944 7.2871 0.0000 35.0000
IMP 17.6307 6.2754 0.0000 35.0000
ASN 17.5410 7.4265 0.0000 35.0000
QUE 17.4169 7.1501 0.0000 35.0000
POW 17.6234 6.7130 0.0000 35.0000
PRO 17.1214 6.9842 0.0000 35.0000
BEL 17.2952 6.7044 0.0000 35.0000
IND 17.8821 5.8345 0.0000 35.0000
REC 17.6166 6.7729 0.0000 35.0000
DTN 17.4786 6.6286 0.0000 35.0000
Table 2. Results of the study

5.2 The Studied Population

Thanks to a partnership with the APEC, we get access to the data collected
by this national organization. It means a large sample of people having passed
the behavioral assessment: 2788 people. In our case, we have restricted our
study to a particular cross-section of population: the experienced executives
who are 45 years old and more and seeking a job. This restriction stems from
a need of the APEC to get a more specific analysis on this particular part of
the population. Indeed, the average behavioral profile of this sample may be
different from the one of the overall population. In our case, we get a cross-
section of the population that contains 613 assessments (one assessment per
subject), in other words 20% of the global sample. This may be interesting for
characterizing some specificities of this population in respect to the reference
308 S. Daviet et al.

population, and then inside this population between some particular profiles
typical of some subgroups of the population.
To study the data in CHIC, we have chosen to transform it into binary
data. In the previous validation study, the computation was made on the
20 traits valued from 0 to 35. In this study, we have used the 10 bipolar
behavioral dimensions discretized in +, 0, − (i.e. for instance: EXT-, EXT0
and EXT+). We have then transform this data into binary data as usually
done in this type of case (i.e. 1 if the characteristic is present, 0 if not).
A sample illustration is shown in Tab. 5.2. The first reason is that CHIC and
the SIA were initially designed to study this type of binary data, and it so
also seems to be the simplest way. The second and most important reason is
that we want to make some indicators appear, in other terms some factors (or
combinations of multiple dimensions). But these indicators may be big trends
more than precise values, because it is more likely to give some consistent
and meaningful classes than discreet values between 0 and 35. That is why
we have driven our study on discretized values.

Ind1 Ind2 Ind3 Ind4 Ind5 . . .

EXT- 0 0 1 0 0
EXT0 1 0 0 0 1
EXT+ 0 1 0 1 0
ASN- 1 0 1 0 0
ASN0 0 0 0 1 0
ASN+ 0 1 0 0 1
...
Table 3. Transformation to binary data

Finally, the data is anonymized, but we also have further information on

those people: their gender, their age, their (previous) activity, etc. It could be
interesting in a second step to use this data to refine our study, but now we
have not used them in this contribution.

6 Why We Used the Statistical Implicative Analysis

and CHIC
For this study, we have used CHIC and the Statistical Implicative Analysis for
multiple reasons. The first one is the ability of CHIC to handle the primary
functions of classical statistics. This is very useful because the first steps of
our study are really simple and classic. Being able to complete this steps and
the rest of the study with the same tool is valuable. The second reason is that
we cannot just rely on a classical statistical study. CHIC gives advanced tools
to study data and can perform hierarchical analysis of data. In our case, it is
 Elaborating Behavioral Referentials with SIA 309

crucial to dichotomize the population to isolate interesting subgroups. Those

groups can then be analyzed by the expert according to their descriptive fac-
tors and strongly support the building of a referential. CHIC also provides at
once similarities, implicative and cohesitive analyses. Those 3 types of analy-
ses are fully complementary. Finally, the tool is visual and quite simple to use.
This is a strong advantage for the expert and the dialog with him. Then, if
we want re-use this analysis process in the future, CHIC is easy to use and
can be used by the expert himself with a short explanation.

7 Global Study
7.1 Goal
We have firstly dealt with the whole studied population of experienced exec-
utives and driven our research along this axis. First of all, we have made a
comparative study between the studied population and the reference popula-
tion thanks to the classical statistical measure: mean. Our goal is to highlight
some relevant indicators that differentiate those executives from the common
individual. Then, we have more deeply studied the inner characteristics of
this population with the tools made available with CHIC. Our goal here is to
find out some specific significant subpopulations both in the statistical and
semantical meaning, and with the expert support, to isolate the behavioral
dimensions implied in this dichotomy and their explanations.
This first step of our study will also help us for the second step. The sub-
population found in this first global study will be more locally analyzed in a
second phase. It will be the basis of the second study to reveal indicators. The
goal is to complete the characterization realized in the global study and con-
firm the first draft of the indicators with a set of complementary dimensions.

7.2 Study of Deviations

Occurrence and frequency (also called mean in CHIC) give apposite information
so that the expert characterizes the studied sample. The standard deviation
has not been used in this first step of the study because it is not meaningful
with binary variables. With these first measures, we have established the most
marked dimensions comparatively to the standard profile (available in the
appendix A). In Tab. 7.2, ASN+, ASN0, COM+ and POW+ (highlighted in
the table) are significantly more important for this population of executives
than for ordinary people. In the light of those results, the significance of ASN+
and POW+ confirms the a priori knowledge of the expert. Indeed, Assertion
and Motivation for power are known features of experienced executives. The
importance of Communication also matches what was expected by the expert.
This first step is not sufficient because the information brought by this
analysis is really poor. The conclusion that an experienced executive is charac-
terized by strong affirmation, communication and motivation for power would
310 S. Daviet et al.

Occurence Frequency Ordinary frequency Absolute gap

EXT- 138 0.23 0.23 0.00
EXT0 254 0.41 0.54 0.13
EXT+ 221 0.36 0.23 0.13
COM- 101 0.16 0.22 0.06
COM0 266 0.43 0.55 0.12
COM+ 246 0.40 0.23 0.17
ANX- 206 0.34 0.23 0.11
ANX0 281 0.46 0.54 0.08
ANX+ 126 0.21 0.23 0.02
ACH- 132 0.22 0.23 0.01
ACH0 265 0.43 0.54 0.11
ACH+ 216 0.35 0.23 0.12
InD- 129 0.21 0.22 0.01
InD0 302 0.49 0.55 0.06
InD+ 182 0.30 0.22 0.08
RIG- 177 0.29 0.23 0.06
RIG0 292 0.48 0.55 0.07
RIG+ 144 0.23 0.22 0.01
ASN- 107 0.17 0.24 0.07
ASN0 217 0.35 0.53 0.18
ASN+ 289 0.47 0.23 0.24
POW- 110 0.18 0.23 0.05
POW0 273 0.45 0.55 0.10
POW+ 230 0.38 0.23 0.15
BEL- 218 0.36 0.23 0.13
BEL0 264 0.43 0.54 0.11
BEL+ 131 0.21 0.22 0.01
REC- 221 0.36 0.23 0.13
REC0 277 0.45 0.53 0.08
REC+ 115 0.19 0.24 0.05
Table 4. Classical statistical measures

be overhasty. Nothing indicates that these characteristics does not split the
global population into two subpopulations or more. To delve into this analysis,
we have completed this study with a more advanced tool of CHIC: similarity
trees.

7.3 Analysis with Similarity Trees

We have used similarity trees with the entropic implication and the Poisson
distribution. Indeed, we have a population of more than one hundred people
and the classical method is not recommended because less discriminatory.
With the same restrictive goal, we have chosen the Poisson distribution.
This second analysis with similarity trees provides a way to determine the
essential classes that partition our population of executives. As we can see on
 Elaborating Behavioral Referentials with SIA 311

Fig. 4, this analysis confirms the significance level of the ASN+ and POW+
dimensions. Indeed, the pair (ASN+, POW+) forms the first significant node
of the tree (marked in bold) with a similarity coefficient of 0.954724. The
dimension EXT+, combined with the (ASN+, POW+) pair, appears to be
relatively significant and discriminative of the population of senior executives.
Those three dimensions underlie most of the significant nodes (similarity of
(EXT+ (ASN+, POW+)) = 0.876296). Therefore, this triplet could be con-
sidered as a good candidate to partition our population and to be a relevant
behavioral indicator.

Fig. 4. Similarity tree

The COM+ dimension appears to be also significant because it forms

the second level node of the tree (similarity = 0.927691). But this node is not
marked as a significant one. Moreover, contrary to the triplet (EXT+ (ASN+,
POW+)), COM+ is not part of a strong partition.
This second analysis has confirmed the weight of ASN+ and POW+. More-
over, we can distinguish three classes where ASN and its three modalities
(ASN-, ASN0 and ASN+) are discriminative. Then, we have categorized three
groups corresponding to each of these modalities. According to the expert,
the ASN- seems to be a defeat factor for the re-employment of executives.
Therefore, the ASN- class may be an interesting basis for building a relevant
indicator showing the ability of an experienced executive to be reemployed.
312 S. Daviet et al.

8 Local Study

8.1 Data Studied and Goal

On the basis of the results discovered through similarity trees, we have studied
each of the three discovered subclasses, discriminated with the ASN dimen-
sion. Indeed, ASN is not sufficient to build a relevant indicator. We need more
clues on what are the main trends of each of these three groups. Thus, we have
used implicative and cohesitive trees to detect link between dimensions. Those
links could then be used to build our indicators as a combination of multiple
dimensions. In the following studies, we have discarded the ASN dimension
that is no more discriminative for each of the 3 subpopulations: ASN-, ASN0
and ASN+. We will only present the statistical details of the study concerning
the ASN- subpopulation, and our conclusions on the 2 other populations. The
conclusions are based both on the comparison between the subpopulations
and the global population of executives, and between the subpopulations and
the ordinary population.

8.2 Analysis of the Subpopulations

ASN- Subpopulation

This subpopulation contains 107 individuals (18% of the global population

of executives). The analysis of the frequencies Tab. 8.2 shows an important
offset of this population compared to the original population of experts for
the following dimensions: POW-, EXT-, ANX+ and RIG+. For the psychol-
ogist, POW- and EXT- are revealing a debasement of self-image and ANX+
and RIG+ indicate an attempt to balance a feeling of insecurity with extra
planning and organisation.
Considering the implicative graph Fig. 5, we can see the two pairs (ACH+,
InD-) and (InD-, ANX+) may be particulary interesting to contribute to a
relevant indicator. But we cannot yet conclude for the (ACH+, InD-, ANX+)
triplet. We need to study the cohesitive tree to know if we can really group
those three dimensions.
The cohesitive tree Fig. 6 and its values Tab. 8.2 give us valuable informa-
tion on the set of dimensions that can be eligible to relevant indicators. By
combining those results with the previous ones, we see that the dimensions
found in the first step (POW-, EXT-, ANX+ and RIG+) are not all eligible
for building an indicator. Indeed, POW- appears at the fifteenth level in a
group with a cohesion value equal to 0.247. Moreover, it can hardly be com-
bined with EXT- as presumed with a simple statistical analysis: the ((BEL0
EXT-) POW-) has a very low cohesitive level.
We can see here the benefit of using an advanced statistical method like
the Statistical Implicative Analysis to prevent eroneous analysis. The POW-
dimension may be, according to the expert, interesting on its own because
 Elaborating Behavioral Referentials with SIA 313

Occurence Frequency Executives frequency Ordinary frequency

EXT- 75 0.70 0.23 0.23
EXT0 32 0.30 0.41 0.54
EXT+ 0 0 0.36 0.23
COM- 41 0.38 0.16 0.22
COM0 52 0.49 0.43 0.55
COM+ 14 0.13 0.40 0.23
ANX- 1 0.01 0.34 0.23
ANX0 27 0.25 0.46 0.54
ANX+ 79 0.74 0.21 0.23
ACH- 17 0.16 0.22 0.23
ACH0 54 0.50 0.43 0.54
ACH+ 36 0.34 0.35 0.23
InD- 58 0.54 0.21 0.22
InD0 43 0.40 0.49 0.55
InD+ 6 0.06 0.30 0.22
RIG- 2 0.02 0.29 0.23
RIG0 30 0.28 0.48 0.55
RIG+ 75 0.70 0.23 0.22
POW- 82 0.77 0.18 0.23
POW0 25 0.23 0.45 0.55
POW+ 0 0 0.38 0.23
BEL- 39 0.36 0.36 0.23
BEL0 42 0.39 0.43 0.54
BEL+ 26 0.24 0.21 0.22
REC- 17 0.16 0.36 0.23
REC0 49 0.46 0.45 0.53
REC+ 41 0.38 0.19 0.24
Table 5. Classical statistical measures and gap

Fig. 5. ASN- implicative graph

314 S. Daviet et al.

Fig. 6. ASN- cohesitive graph

Levels Cohesion
1 (REC- BEL-) 0.993
2 (COM+(REC- BEL-)) 0.991
3 (RIG- InD+) 0.966
4 (BEL+ REC+) 0.962
5 (ACH+ InD-) 0.948
6 ((RIG- InD+)ACH-) 0.944
7 ((ACH+ InD-)ANX+) 0.939
8 (REC0 COM0) 0.893
9 (ANX- RIG0) 0.876
10 ((BEL+ REC+)COM-) 0.868
11 ((COM+(REC- BEL-))RIG+) 0.856
12 (ANX0 InD0) 0.559
13 (((BEL+ REC+)COM-)ACH0) 0.366
14 (BEL0 EXT-) 0.313
15 ((BEL0 EXT-)POW-) 0.247
16 (POW0 EXT0) 0.155
Table 6. Cohesitive values
 Elaborating Behavioral Referentials with SIA 315

it reveals the loss of leadership and can explain the difficulty of this sub-
population to reintegrate the working world. However, the expert has found
interesting those combinations of dimensions:
• (((BEL+ REC+) COM-) ACH0): relies on others by accepting concessions,
so as to lighten his/her work load,
• ((COM+ (REC- BEL-)) RIG+): hides behind an inflexible, aloof and even
strongly opposed behavior.
In the light of those results for the ASN- subpopulation, the expert has char-
acterized a set of dimensions that is meaningful according to its psychological
knowledge. Hereunder, you can see the indicators built:
• Indicator of adaptation: REC0/COM0 (17% of the sample),
• Indicator of illusion: RIG-/InD+/ACH-,
• Indicator of cry for help: BEL+/REC+/COM-,
• Indicators of autistic withdrawal:
– passive: EXT-/POW-/BEL0,
– offensive: COM+/REC-/BEL-,
• Indicator of strictness by:
– obstinacy: ACH+/RIG+,
– nervous tensing up: ACH+/InD-/ANX+.
If we consider the implicative graph, we can see that almost all the combina-
tions over the 0.90 threshold have been kept and construed by the expert.

8.3 ASN0 Subpopulation

This subpopulation contains 217 individuals (35% of the global population

of executives). The study of frequencies shows that this subpopulation has a
highly lower offset with the ordinary population than the previous ASN- sub-
population. The highest offset is on POW0 that is a quite neutral dimension
and, according to the expert, not really significant on its own. Once more, the
classical statistical tools are not sufficient to build our indicators.
Using the SIA, both similarity and cohesitive trees and implicative graphs
show that the (REC-, COM+) pair is a strong indicator of this subpopulation.
According to the expert, it can reveal a reject of others. Again, almost all
the combinations discovered by the implicative graph have been kept and
interpreted by the expert. However, we had to lower the thresholds to 0.80,
because this subpopulation is poorly characterized. Hereunder, you can see
the basis of indicators built by the psychologist expert:
• Indicator of influence: POW+/ANX0 and POW+/RIG0,
• Indicator of open-mindedness as:
– curiosity: InD+/ANX-/POW0,
– intellectual adaptability: InD+/RIG-/POW0,
– cleverness: InD+/RIG-/ACH-,
316 S. Daviet et al.

• Indicator of interpersonal skill: REC0/COM0 by:

– sharing: REC0/COM0/POW-,
– compliance: REC+/COM-/POW0,
– conviviality: REC+/BEL+/POW0,
– benevolence: BEL+/REC+/COM-,
• Indicator of autistic withdrawal: EXT-/BEL-,
• Indicator of reject of others: REC-/BEL-/COM+,
• Indicator of strictness and nervous tensing up: ANX+/RIG+/InD-/ACH+.

8.4 ASN+ Subpopulation

This subpopulation contains 289 individuals (47% of the global population of

executives). This subpopulation is better characterized than the ASN0 popu-
lation. Indeed, according to the study of frequencies, the offset is well marked
on multiple dimensions: EXT+, COM+, ANX-, RIG- and POW0. Those five
dimensions seem to validate the apriori knowledge of the expert that the
ASN+ population is the best able to return to the working world. Represent-
ing almost fifty percent of the overall executive population, this subpopulation
follows pretty much the same trend, and the classical statistical tools are not
sufficient to get relevant specific information. As usual, CHIC has been used
to go into detail.
The SIA has brought new information on combination of interesting di-
mensions. This subpopulation is well characterized and many indicators have
been found:
• Indicator of enthusiasm: ACH-/RIG-/InD+/ANX-/EXT+/POW+,
• Indicator of interpersonal skill by:
– involvement: BEL0/REC0/COM0,
– conviviality: BEL+/REC+/COM-,
• Indicator of vigilance: ANX+/EXT0,
• Indicator of strictness:
– interpersonal: EXT-/BEL-/REC-/COM+,
– intellectual: InD-/ACH+,
– organizational: RIG+/POW0.

9 Results and Outlooks

This study has led to interesting discoveries according to the psychologist ex-
pert. The method previously used was designed only over single dimensions
with classical statistical tools. As our study has shown, this could give impre-
cise results, and in some particular case, erroneous ones. Moreover, it is not
the results but the expert’s interpretation that can be biased by statistical in-
formation that do not match those he/she expects to be working with. Thus,
this study has proven the interest of using more advanced statistical tools and
 Elaborating Behavioral Referentials with SIA 317

data analysis methods in the field of psychology, more accustomed to classical

statistics.
Thanks to this study, the expert has been able to build a set of relevant
indicators on our initial population of experienced executives seeking a job. He
has identified three main groups on the assertion (ASN) dimension. On each
group, numerous indicators have been designed by the expert. Because it was
a prospective study, the results have then been confronted with other data
of the APEC on the studied subjects: it appears that the indicators found
were quite relevant when looking at the global behavior of each of these three
groups. Indeed, each of these groups has a particular behavior considering its
working world reintegration. The group in weak assertion is mainly charac-
terized by a longer reintegration period and by a feeling of defeat relative to
their unemployment situation, whereas the group in strong assertion has a
higher rate of success in reemployment and shows a more positive behavior
towards their situation. The group in medium assertion is less clearly defined
than the other two and the behavior of its subjects is less uniform. Some of
them follow the trend of the ASN- group, others that of the ASN+. The indi-
cators built for each group match those observations. Most of the indicators of
the ASN- population (indicators of illusion, cry for help, autistic withdrawal,
strictness) can be considered as vigilance points denoting potential difficulties
for some people of this population. This does not mean that a person of the
ASN- subpopulation is bound to fail, it means that people of this group must
be watched more closely. This is exactly the goal of the APEC and it is also
a success of our study. Indeed, it shows that CHIC can be used as a decision
support tool, combined with the psychological assessment tool Echo.
The results of this study are really hopeful. First, we have been able to
show the coherence of the PerformanSe Echo tool over the studied population:
the experienced executives, the study of deviations match the knowledge of
the expert. Then, we have been able to dichotomize three characteristic sub-
populations and highlight the meaningful collusions of several dimensions on
which to build our indicators with the expert’s support. Third, it seems that
the Statistical Implicative Analysis is a means for semi automatically building
behavioral referentials. From a study on a large sample of behavioral assess-
ments, we have built a decision support tool that can be used for counseling,
training or access to a job. For the moment, the study is still in progress
and we trying to take advantage of supplementary variables available in the
sample. We are also thinking of studying the historical evolution of these pop-
ulations thanks to Statistical Implicative Analysis or other multi-dimensional
data mining methods.

References
1. R. Agrawal and R. Srikant. Fast algorithms for mining association rules. In
J.B. Bocca, M. Jarke, and C. Zaniolo, editors, 20th International Conference on
318 S. Daviet et al.

Very Large Data Bases, VLDB’94, pages 487–499. Morgan Kaufmann, 1994.
2. J. G. Carlson. Recent assessment of the mbti. Journal of Personality Assess-
ment, 49(4), 1985.
3. M. Carlyn. An assessment of the myers-briggs type indicator. Joumal of Per-
sonality Assessment, 41:461–473, 1977.
4. R. Couturier. Traitement de l’analyse statistique implicative dans chic. In
Journées sur la fouille des données par la méthode d’analyse implicative, pages
33–55, 2001.
5. D. Cowan. An alternative to the dichotomous interpretation of jung’s psycholog-
ical functions: Developing more sensitive measurement technology. In Journal
of Personality Assessment, volume 53, pages 459–471, 1989.
6. U. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy, editors. Ad-
vances in Knowledge Discovery and Data Mining. MIT Press, Cambridge, MA,
1996.
7. J. George and G. Jones. Organizational Behavior. Prentic Hall, Upper Saddler
River, NJ, 3rd ed. 2004 edition, 2002.
8. L. R. Goldberg. Language and individual differences: The search for universals
in personality lexicons. Review of Personality and Social Psychology, 2:141–165,
1981.
9. R. Gras. L’implication statistique : une nouvelle méthode exploratoire de don-
nées. La Pensée sauvage. 1996.
10. R. Gras, H. Briand, P. Peter, and J. Philippe. Implicative statistical analysis.
In Proceedings of International Congress I.F.C.S., Kobe, Tokyo, 1997. Springer-
Verlag.
11. R. Harvey, W. Murry, and S. Markham. Evaluation of three-short-form versions
of the mbti. Journal of Personality Assessment, 63(1):181–184, 1994.
12. J.H. Johnson and T.A. Williams. Using a microcomputer for on-line psycholog-
ical assessement. Behavior Research Methods & Instrumentation, 10:576–578,
1978.
13. I.C. Lerman. Classification et analyse ordinale des données. Dunod, 1981.
14. I. Myers. The myers-briggs type indicator. Educational Testing Service, 1962.
15. P. Myers. Gifts Differing. Understanding Personality Type. Davies-Black Pub-
lishing, 1995.
16. T. Patel. Comparing the usefulness of conventional and recent personality as-
sessment tools: Playing the right music with the wrong instrument? Global
Business Review, 7(2):195–218, 2006.
17. V. Philippé, S. Baquedano, R. Gras, P. Peter, J. Juhel, P. Vrignaud, and
Y. Forner. étude de validation : Performanse echo, performanse oriente. Tech-
nical report, Study realized with the collaboration of PerformanSe, Laboratoire
COD de l’École Polytechniquede l’Université de Nantes, Laboratoire de Psy-
chologie Différentielle de l’Université de Rennes 2, 2004.
18. B. S. Stein and J. D. Bransford. Constraints on effective elaboration: effects of
precision and subject generation. Journal of Verbal Learning andVerbal Behav-
ior, 18:769–777, 1979.
19. O. Tzeng, D. Outcalt, S. Boyer, R. Ware, and D. Landis. Item validity of the
mbti. Journal of Personality Assessment, 48(3), 1984.
20. T. Vacha-Haase and B. Thompson. Alternative ways of measuring counselees’
jungian psychological-type preferences. Journal of Counselling and Develop-
ment, 80, 2002.
 Elaborating Behavioral Referentials with SIA 319

21. J. S. Wiggins, editor. The Five-Factor Model of Personality: Theoretical Per-

spectives. Guilford, New York, 1996.
Fictitious Pupils and Implicative Analysis:
a Case Study

Pilar Orús and Pablo Gregori

Universitat Jaume I, Castellón E-12071, Spain

{orus, gregori}@mat.uji.es

Summary. We present a case study, in the context of Didactics of Mathematics, in

which we adopt the methodology of using fictitious data in the Statistical Implica-
tive Analysis. On the one hand, unlike supplementary variables, the fact of adding
fictitious data to the sample does modify analyses results, so caution is needed.
On the other hand, fictitious students are a tool for better understanding the data
structure resulting from the analyses.

Key words: Contribution, entropic implication, fictitious subject, intensity of im-

plication, quasi-implication, statistical implicative analysis, typicality.

1 Introduction
The use of multivariate analysis in the field of Didactics of Mathematics (DM)
has already got a long tradition in the frame of fundamental didactics. Im-
portant references can be found among the contributions of Journées de Caen
(1995 & 2000), such as [3] and [10], in which new statistical tools are provided,
motivated in the context of DM as in many other occasions, but fruitful for
both the fields of DM and multivariate statistics.
In the usual multivariate methods (generally factor analysis and princi-
pal component analysis), Brousseau [3] uses some supplementary individuals
(fictitious individuals) in his data in order to be able to compare the a pri-
ori and a posteriori analysis of a questionnaire. The a priori analysis of the
questionnaire leads to certain criteria of characterization of its questions (the
variables). In this way, two matrices are obtained: one coming from the pre-
experimental analysis (the a priori matrix of the questionnaire: criteria ×
questions) and the empirical matrix, made of the collected data, where ques-
tions remain characterised by the present sample (answers × questions).
Fictitious individuals allow for the simultaneous consideration of both
the pre-experimental criteria and those provided by the sample in a single
matrix. Fictitious individuals, as features of the variables involved in the
P. Orús and P. Gregori: Fictitious Pupils and Implicative Analysis: a Case Study, Studies in
Computational Intelligence (SCI) 127, 321–345 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
322 Pilar Orús and Pablo Gregori

experiment, contribute both to the improvement of the knowledge on the

variables —enhancing the information management furnished by the sample
and thoroughly analysed a priori —, and to the comparison of the a priori
and a posteriori behavior of the same variables.
The study of dependences between pieces of knowledge, approached in [9],
highlighted the existence of non symmetric relations between variables —in
a context of didactics— and motivated the search of new tools to analise
such relations. Therefore the Statistical Implicative Analysis (SIA) is intro-
duced [10–12] and mainly developed in research in DM [4,13,15,19], although
it has also been used in other fields, such as in [7]. This theory is located
among other procedures of data mining [1, 20, 21].
Our contribution intends to be located in the intersection of both con-
texts. On the one hand, we support the didactical interest of the fictitious
individuals in the multivariate analysis of data as Brousseau has done in [3].
On the other hand, we consider SIA as a very appropriate and powerful tool
for the processing of data obtained in the DM research. In this frame, the
contribution [18] meant an empirical and naive approach to the use of fic-
titious individuals within SIA, following the direction of [22], that explored
the contrast between the a priori analysis of a didactical situation and the
contingence found in the experimentation.
Contribution [18] was an exercise of experimentation and observation,
where the notion of fictitious individual was used just as any other resource
in the several data analysis provided by the SIA, by means of the statistical
software CHIC (acronym of Classification Hiérarchique Implicative et Cohési-
tive, see [5]). The role we intended to assign to fictitious students was similar
to the one played by supplementary variables with respect to ordinary vari-
ables. That is, to use their relative placement in the final structure of data,
but leaving the computation of ordinary variables untouched [6]. Obviously
the nature of SIA —and its implementation in CHIC — do not offer this pos-
sibility, unlike other techniques such as factor analysis or principal component
analysis [3].
Following the classical theory and using the Poisson distribution, several
implicative analysis were performed on a binary valued matrix containing the
results of a Mathematics test taken by 690 first-year students of University
Jaume I (Spain). Five profiles were also used to classify the items a priori,
and were considered as supplementary students, with their answers to each
item forming the a priori matrix of the test.
The presence of these five fictitious students in procedures such as the typi-
cality and contribution of individuals to the formation of similarity, implicative
and cohesion classes of items show their potential use in the didactical inter-
pretation of results in [18]. But new questions concerning this methodology
arise, such as the ones regarding the conditions under which their use leads
to reliable conclusions.
This work shows the instrumental role of fictitious students in didactical
analysis performed through SIA, which could be transferred to other domains
 Fictitious pupils and implicative analysis 323

of application, according to the available data, by the consideration of ficti-

tious extra data.

2 Case Study Description

2.1 Aim of the Study

The aim of this chapter is the search of new potential uses of SIA, within the
context of DM, in order to promote the development of this theory and its
applications. In this way, we present and study a conjecture on the introduc-
tion of fictitious individuals in the sample data matrix and in the subsequent
processing through CHIC [5]. On the one hand, this procedure offers assis-
tance in the didactical interpretation of the results. On the other hand, the
final results of the analysis are perturbed by this artificial data, so that a non
negligible size of this perturbation would invalidate the advantages achieved
with respect to the interpretation. Therefore, this work evolves in two direc-
tions: one in DM, and the other one in the methodology of SIA, as a corollary
of the first one.
We show, in the several analyses performed, not only results but also the
philosophy concerning the use of fictitious students: when, where and what
they are introduced for. Firstly, we present data. Then, in Sect. 3, we show
the results of the different SIA procedures applied to data under the classical
version [18], with and without fictitious data, keeping track on the differences,
checking that they are reasonably small, and stressing on the gain of informa-
tion through the procedure. Next, in Sect. 4, we take profit of the conclusions
of the previous section to design a new set of fictitious students with which
we analyse the structure of the same dataset (using the entropic version of
the SIA), improving the obtained information, as well as the appreciated dif-
ferences with respect to the previous results (with classical implication).
Having posed the question on how the structure of SIA results do vary
when adding a small number of individuals to an existing sample, we show,
in Appendix A, and for the interested reader, a short introduction to quasi-
implications and their intensity, and a result on its variation under addition
of new individuals to the sample. In the case study shown in this chapter, the
number of new individuals are less than 1% of the sample size.
Finally, the questionnaire used to obtain the data of the study is shown in
Appendix B.

2.2 Description of the Case

A test on the initial skills in Mathematics has been conducted over the popu-
lation of first-year students of University Jaume I (UJI) of Castellón (Spain)
since 2001 [14, 17, 19]. For the first time, it was part of the development of a
DM research project of Bosch and collaborators (see [4]) and a PhD thesis [8].
324 Pilar Orús and Pablo Gregori

The items of the test were selected in order to ascertain some given didac-
tics hypotheses on the didactical discontinuities between the mathematics at
the pre-university and university levels. Adapted versions of that test have
been conducted in the subsequent years, being part of several Educational
Improvement Projects promoted by the institution. They have been used to
help the UJI Mathematics Department professors to get to know the skills of
the students they are going to work with. As a consequence,they would assure
what is expected for them to master, and it would allow them to freely modify
their didactical strategies of education.

Data.

The present study is based on data obtained through the experimentation

of the questionnaire (test) on the initial skills on mathematics of first-year
students at University Jaume I, more precisely belonging to the High School
of Technology and Experimental Sciences, at the beginning of the academic
year 2003–04.
The questionnaire consists of 17 questions corresponding to 21 single items
whose answers are coded as 0 (fail or unanswered) or 1 (success). Then we
manage 21 binary variables labelled as P1, P2, P3, P4, P5, P6a, P6b, P7,
P8, P9, P10, P11, P12, P13a, P13b, P14a, P14b, P15, P16, P17a and P17b.
Additional supplementary variables have been taken into account, such as the
kind of degree, attendance of mathematics preparation lectures, origin of the
student and result in the State University Access Test, adding up to 16 of
them. Contribution [18] focused on the role of these variables. In the present
work, we manage a Boolean contingence table of 21 variables and 690 students.

The a priori matrix (MAP) of the questionnaire: fictitious

students.

A first a priori classification of the items in the questionnaire, according to

their type of knowledge and the task involved is shown in Table 1.

Type of knowl- Type of task Item

edge
problem (p) P1, P6a, P6b, P8, P10, P12, P13a,
Algebra (A) P13b, P15, P16, P17b
graphical (g) P17a
exercise (e) P2, P7
problem (p) P8, P14b, P17b
Calculus (C) graphical (g) P3, P17a
exercise (e) P3, P4, P5, P9, P11, P14a
Table 1. Type of knowledge and task involved in the items of the questionnaire.
 Fictitious pupils and implicative analysis 325

The item features in Table 1 have been used in our first SIA analyses,
under the classical theory, in order to define fictitious students. Let us note
that this classification is uneven in the sense that the cardinals of classes are
very different. Fictitious students are Algebra (A) and Calculus (C), scoring 1
only in items classified under those respective type of knowledge, and problem
(p), graphical (g) and exercise (e), scoring 1 only in items classified under
those respective types of task.

3 Application of SIA using the Classical Implication

In this section we present our methodology, based in the analysis of the optimal
groups of individuals regarding the contribution to the formation of classes
and their typicality, in order to improve the interpretation of the rules and
the quantity of information about the sample. We shall present what will be
shown in particular cases, without the intention to cover all aspects.

3.1 Classification Analysis

The original data processed through software CHIC 3.7 [5] leads to a clas-
sification tree, after the similarity index defined by I.C. Lerman [16]. The
classification tree for the data containing the supplementary fictitious stu-
dents is so close to the one for the original data, that eye inspection cannot
distinguish among them. This fact allows us to go beyond the analysis re-
sult, examining the role of new students (item features) in the constitution of
classes and rules.
If we relabel items, reflecting the criteria of the a priori analysis, and
details on the type of task, we display the dependence on the variable class
formation in a better way, and we can then compare it to the fictitious stu-
dents appearing in the optimal groups of classes (see Fig. 1). Codes ‘m’, ‘t’,
‘i’ do represent mathematical modelisation (m), algorithmic technique (t), in-
terpretation or judgment (i). The relation between former and new labeling
is expressed in Table 2.
Fictitious students taking part in optimal groups for the contribution to
the formation of classes as well as for typicality are actually the same (however
we display it in Fig. 1 only for significant knots).
Classification tree shows significant knots at levels 1, 6, 9, 12, 15, 17, et
19, being level 12 the most significant.
For instance, items P4, P5 of class 1 are, a priori, calculus exercises, and
those features do appear as fictitious students in the optimal group of indi-
viduals contributing to the formation of that class. On the other side, level 9
items (((P15, P17a), P16), P17b) are algebra problems, but only the fictitious
student Algebra do appear at the contribution optimal group of students.
326 Pilar Orús and Pablo Gregori

Item Relabeled as Meaning

P1 1Apt Algepra problem involving an algorithmical technique
P2 2Aet Algebra exercise involving an algorithmical technique
P3 3Cgt Calculus graphical representation involving an algorithmical
technique
P4 4Cet Calculus exercise involving an algorithmical technique
P5 5Cet Calculus exercise involving an algorithmical technique
P6a a6Apj Algebra problem focused on the interpretation of results
P6b b6Apj Algebra problem focused on the interpretation of results
P7 7Aet Algebra exercise involving an algorithmical technique
P8 8Xpm Algebra and Calculus problem requiring modelisation
P9 9Cet Calculus exercise involving an algorithmical technique
P10 10Apj Algebra problem focused on the interpretation of a situation
P11 11Cet Calculus exercise involving an algorithmical technique
P12 12Apj Algebra problem focused on the interpretation of results
P13a a13Apm Algebra problem requiring modelisation
P13b b13Apm Algebra problem requiring modelisation
P14a a14Cet Calculus exercise involving an algorithmical technique
P14b b14Cpm Calculus problem requiring modelisation
P15 15Apm Algebra problem requiring modelisation
P16 16Apgm Algebra problem with graphical representation requiring mod-
elisation
P17a a17Xgt Algebra and Calculus graphical representation involving an
algorithmical technique
P17b b17Xpm Algebra and Calculus problem requiring modelisation
Table 2. Relabeling of test items: Each item is described as (1) item number (now
with prefix a or b if it had that suffix in the initial labeling) (2) series of letters
expressing features of the item, (A: Algebra, C: Calculus, X: both Algebra and Cal-
culus, e: exercise, p: problem, g: graphic, m: mathematical modelisation, t: algorithmic
technique, j: interpretation or judgment).

This kind of arguments can be drawn at each class formation. To sum-

marise, the analysis of the contributions of fictitious students in the forma-
tion of the classes in the similarity tree shows that, in a first step, the types
of knowledge (Algebra and Calculus) arise as determinants in the class forma-
tions whereas the types of task arise associated to type of knowledge (exercise
attached to Calculus and problem attached to Algebra).

3.2 Implicative Analysis

The results of the implicative analysis driven through CHIC on our question-
naire, using the classical implication and the Poisson distribution, depict rules
among the questions coming from the answers given by the students sample.
Figure 2 (left and center) represents those quasi-implication rules using, re-
spectively, the real sample and the enlarged one (with fictitious students).
 Fictitious pupils and implicative analysis 327

m
m

m
m
m
t

t
e

p
p

A ,
m
t

j
C

e
A

X
a

p
t

t
t

t
C
g

A
p

e
4

7
C

C
A

X
1

6
1

b
C+e A+p
(C+e)
C
(C) A+p A
(A)
C+A A+p
(C+A)

A C +A
(A) (C+A) C+A
(C+A)

Fig. 1. Similarity tree of the questionnaire using new fictitious students (C: Calculus,
A: Algebra, e: exercise, p: problem, g: graphic) and highlighting them when belonging
to optimal groups regarding the contribution and typicality (between parenthesis).
Items are labeled as shown in Table 2.

If we focus on the implicative graph regarding items belonging to class

12, we can perceive three subgraphs A, B and C (see Fig. 2, right). They
represent rules among the items that constitute themselves complete classes
at significant knots 1, 6 and 9, respectively. The implication P4→P5 (between
calculus exercises) in subgraph A, show that students who perform correctly
the derivative of a rather complex function (f (x) = 5/(3x−2)2 ), also generally
R3
do perform correctly a definite integral of a very simple function ( 1 2axdx).
The rules among algebra problems (P17a → P17b → P15 → P16) of
subgraph C, specifies a relation between translation activities in two registers:
a graphical one and an algebraic one.
The subgraph B gathers the different levels of aggregation conforming
class 6. Rule P9 → P14a points out the relation between two calculus exer-
cises concerning limits. The chain (P14b → P14a →P13b → P13a) rapports
two function problems, and what we can observe is the following: firstly, the
performance of the b-part of a problem implies the performance of the a-part,
and secondly, success in problem P14 (involving an exponential function) im-
plies success in problem P13 (involving an affine function). Therefore, B spec-
ifies rules between abilities and more complex knowledge, i.e., non algorithmic
knowledge.
Analysing the role of fictitious students in the contribution to the formation
of rules, we find that problem and Algebra play it in (a17Xgt → b17Xpm →
15Apm → 16Apgm) and (b14Cpm → a14Cet → b13Apm → a13Apm →
15Apm → 16Apgm). On the other side exercise and Calculus play it in (9Cet
→ a14Cet → 5Cet) and (4Cet → 5Cet). Similarly, only Calculus appear in
328 Pilar Orús and Pablo Gregori

9Cet b14Cpm

p9 p14b p9 p14b B
a17Xgt a14Cet

p11 p17a p14a

p17a p14a p11

p4 p13b
4Cet b13Apm
17b p13b p4

A
17b p13a p5

p13a p5 p6b
b17Xpm 5Cet a13Apm

p15 p8 p6b

p8 p15 p7 p6a

p16 p10 p6a p7

15Apm

p10 p16 p2
p12 p3 p2
C

p1 p3 p12 16Apgm
p1

Fig. 2. Implicative graph of data without (left) and with (center) fictitious students.
Implicative graph of the subset of items forming class 12 in the classification tree
(right). Items are labeled as shown in Table 2.

(b14Cpm → a14Cet) and (a17Xgt → 4Cet → 5Cet), and only Algebra in

(b14Cpm → a14Cet → b13Apm → a13Apm) and (a17Xgt → b17Xpm →
15Apm).
We have experienced that the search of fictitious students in contribution
optimal groups of the implication chains of the implicative graph brings on
the features of involved items, completing and tuning the information of the
a priori categorisation, which can in turn fail to explain the observed data.
We confirm again the linked features exercise-Calculus and problem-Algebra.
Even when observing only slight differences between implicative analysis
with and without fictitious students, we don’t think it is appropriate to pro-
ceed with subsequent analysis using fictitious students until some stability
or robustness of implicative analysis results under addition of new data are
proved.

3.3 Cohesion Analysis

The cohesion tree of the original data shows significant knots at levels 1, 3, 8,
14 et 17, being the most significant the one at level 1 (see Fig. 3).
Here we discover again that algebra problems P17a, P17b, P15 and P16
are linked in the same class, but now, the chain of implications given in the
implicative analysis shows new nuances in the cohesion tree: [P17b → (P17a
 Fictitious pupils and implicative analysis 329

a
b

5
6

1
p

p
Fig. 3. Cohesion tree of items in the questionnaire.

→ P16) → p15]. This class puts together, not symmetrically, two translation
tasks, one in graphical language and another one in formal language.
In level 14 we find items from implicative subgraph B. The cohesion analy-
sis shows a stronger relation between a- and b-parts of each problem P13 and
P14, than the implicative analysis did.
Now, the introduction of fictitious students lead to a similar cohesion tree,
where only a new level of aggregation (level 11) is significant. Cohesion is
strong: it equals 1 until level 15 and greater than 0.99 in the lowest significant
level. We display in Fig. 4 the cohesion tree with relabeling of items (see
Table 2), and fictitious students appearing in the contribution and typicality
optimal groups of each significant knot.
If we consider Fig. 4 as the compilation of all the provided information,
we can assess the existence of three classes, T, O and M, being M the most
significant, not only because of the signification of the meta-rule relating all
the items, but also because of the significance of the included rules.
Class M establishes a meta-rule characterised by the need of modelisa-
tion. It comprises complex questions, which require the interpretation of the
context and the knowledge of mathematical models in several frameworks
(graphical, algebraic, functional, . . . ). It establishes a dissimetry among two
meta-rules, also significant and already analysed [P14b →P14a] → [(P13a →
P13b) → P10)], which specifies rules between abilities and high level knowl-
edge (non algorithmic), and [((P17a → (P17b → P16)) → P15], which reveals
relations between graphical and algebraic registers. We have found fictitious
students Algebra and problem in the optimal groups for the contribution to
the formation of classes. Also, the presence of student Calculus in [P14b →
P14a] allows us to formulate that the ability in mathematical modelisation of
330 Pilar Orús and Pablo Gregori

a C m

3 m

b p m

7 m

A m
b A t

1 p t

m
1 p
1 e
1 p
0 p

1 p
6 g
5 g
1 j,
C t
6 pj
C j

1 j

3 pm
4 e

a C
9 p

1 p

b pa

p
1 A

a X
1 X
1 et

5 et
7 et

1 gt
2 et
8 et
C
A
A

A
4
4
3

7
C

C
A
A
X

A
6

1
b
a

C+e
A+p (C + e)
(A + p) A+p
(A + p)
C

C
(C) B
A+p
C+e (A + p)
(C + e) O
f +72 al.
(f) A+e
(A + e)
f +91 al
e
(C)
A
(A+p)
M
T f +218 al
e
A
A (A+p)
(A+p)

Fig. 4. Cohesion tree of the questionnaire using fictitious students (A: Algebra, C:
Calculus, e: exercise, p: problem, g: graphic). We point them out into squares when
they belong to optimal groups regarding the contribution and typicality (the last
one between parenthesis). φ is used to indicate no fictitious student belong to the
optimal group. Items are labeled as shown in Table 2.

problems implies the ability in its resolution technique. And it confirms the
characterisation that we have done to class M, strengthened by the presence
of students Algebra and problem in the optimal group of typicality of the class,
and all its significant subclasses.
Class T groups together practically all items which can be solved with al-
gorithmic techniques. It also includes some rules determined by the presence
of Calculus and exercise as students in the optimal groups for the contribu-
tion. Here the type of knowledge of the item is not relevant but the applied
technique used to solve it.
Class O is determined by the contingency: no fictitious student appear
among the 91 individuals contributing to the formation of this class. Calculus
appears only in the optimal group for the contribution to the rule P8 → P3,
together with 93 more students, but it disappears at the following level.
 Fictitious pupils and implicative analysis 331

3.4 Conclusions of SIA with Classical Implication

Classification and cohesion analyses based on fictitious students bring on com-

plementary information, mainly related to the set of variables (P4, P5, P13a,
P13b, P14a, P14b, P17a, P17b, P16, P15) of our questionnaire. They have
been used as tools for the mining of our dataset, investigating the nature of
the formation of classes to the a priori emphasised features of items. However,
special attention has to be paid to the size of changes that fictitious students
produce over global results.
We should highlight the strong dependence existing between the features
exercise and Calculus (in the questionnaire, exercises were mostly related to
the systematic application of a technique, generally dealing with functions,
and mainly identified with the type of knowledge Calculus) and similarly,
between problem and Algebra. This fact explains they usually appear in pairs.
The inspection of the modifications suffered by SIA results after the intro-
duction of fictitious students has allowed us to detect weakness in our a priori
analysis, motivating a new a priori feature classification of items, then new
fictitious students, and new possibilities of explanations of the SIA analyses.

4 New Fictitious Students and the Application of SIA

with Entropic Implication

We have chosen to use the results of previous hierarchical trees of similarity

and cohesion of variables to design new criteria of classification of items (that
is, new fictitious students, see Table 3), by keeping the cardinal small, 5,
and being aware of the eventual changes that this new a priori matrix could
generate in the structure of original data.
Under the new criteria, item codes are updated as shown in Table 4.
The new 5 selected criteria (F, E, t, i, m) do not represent a disjoint
partition of the questionnaire, but allow us a different way of characterising
items. That is to say, new fictitious students that we shall use in the following
analyses. These analyses also need to use a convenient labeling of items that
keeps the codes of the former classification as problem or exercise, allowing us
to tune our didactical analysis.
In this section we proceed through a new SIA of our data that has been
conducted using CHIC 3.7, choosing the entropic version and the Poisson
distribution.

4.1 Similarity Tree

Taking into account the new fictitious students and the according item labels,
we observe that the similarity tree is practically the same one as the one built
from original data, therefore only one tree is shown in Fig. 5.
332 Pilar Orús and Pablo Gregori

Item
Requires modelisation P8, P10, P12, P13a, P15,
Type of task (m) P17b
Requires interpretation P6a, P6b, P8, P10, P12,
(i) P13b, P14b, P16
Application of technique P1, P2, P3, P4, P5, P7,
(t) P9, P11, P14a, P17a,
P17b
Function P2, P6a, P6b, P8, P10,
Type of knowledge
(F) P12, P15, P16, P17b
Equation P3, P4, P5, P8, P9, P10,
(E) P11, P13a, P13b, P14a,
P14b, P16, P17b
Table 3. Redefinition of types of knowledge and task of items in the questionnaire
(new fictitious students).

Then we use the analysis including the fictitious data in order to get new
conclusions. Three classes are identified, respectively dubbed C1 , C2 and C3 .
Class C1 contains the highest statistical significance levels of aggregation, and
shows a class of items characterised by applicability tasks in particular context
(what is commonly referred as problem and labelled by ‘p’) even though if the
nature of the application varies: use of algorithm (P1, P3, P14a), contextual
interpretation (P10, P13b, P14b), mathematical modelisation (P13a) or a
combination of them (P8). Class C2 may be the result of chance, since it does
not contain significant levels. Finally, class C3 gathers items concerning the
graphical nature of functions or their operations.
The inspection of optimal groups of students for both the contribution and
the typicality leads to the information shown in Fig. 5.
We can state that, in the classification analysis, fictitious students only
contribute to the formation of most of subclasses along the first levels of
aggregation, because of the heterogeneity of the final classes. However, they
belong to typicality optimal groups along all subclasses and classes.

4.2 Cohesion Tree

The cohesion tree of the original data resulting under the entropic theory
(Fig. 6, left) shows a weaker structure: isolated and binary rules are abundant,
and only one meta-rule is present. Nevertheless, it is practically a simple rule,
since the implication of the premise involves two items which are parts ‘a’ and
‘b’ of a double question of the questionnaire. This cohesion tree does not seem
to show relevant information. However, the consideration of fictitious students
(Fig. 6, right) adds up structure to data: it reduces the number of isolated
items whereas the number of meta-rule raises, involving items belonging to
class C1 of similarity tree.
 Fictitious pupils and implicative analysis 333

Item Relabeled as Meaning

P1 1pt Problem involving an algorithmical technique
P2 2Eet Equation-related exercise involving an algorithmical technique
P3 3Fgt Graphical representation of a function involving an algorithmi-
cal technique
P4 4Fet Function-related exercise involving an algorithmical technique
P5 5Fet Function-related exercise involving an algorithmical technique
P6a a6Epi Equation-related problem focused on the interpretation of re-
sults
P6b b6Epi Equation-related problem focused on the interpretation of re-
sults
P7 7et Exercise involving an algorithmical technique
P8 8Xpmi Equation- and function-related problem requiring a modelisa-
tion and the interpretation of results
P9 9Fet Function-related exercise involving an algorithmical technique
P10 10Xpi Equation- and function-related problem focused on the inter-
pretation of a situation
P11 11Fet Function-related exercise involving an algorithmical technique
P12 12Epi Equation-related problem focused on the interpretation of re-
sults
P13a a13Fpm Function-related problem requiring the modelisation of a situ-
ation
P13b b13Fpi Function-related problem requiring the modelisation of a result
P14a a14Fet Function-related exercise involving an algorithmical technique
P14b b14Fpi Function-related problem focused on the interpretation of a
result
P15 15Epm Equation-related problem requiring the modelisation of a situ-
ation
P16 16Xpgi Equation- and function-related problem involving the interpre-
tation of a graphical representation issue
P17a a17Fgt Function-related graphical representation solvable by an algo-
rithmical technique
P17b b17Egt Equation-related of a graphical representation situation solv-
able by an algorithmical technique

Table 4. Relabeling of items following the new fictitious students: Each item is
described as (1) item number (now with prefix a or b if it had that suffix in the
initial labeling) (2) series of letters expressing features of the item, (E: Equation,
F: Function, X: both Equation and Function, e: exercise, p: problem, g: graphic, m:
mathematical modelisation, t: algorithmic technique, i: interpretation or judgment).
334 Pilar Orús and Pablo Gregori

1 pm

x
a pm

7 i
4 t

7 t
4 i

E i

a pg

g
a Fp

b Fe

2 Fp

b Fg
a mi
8 pi

b pi

9 pi

1 pi
1 et

E
F

5 t
1 t

4 t

7 t

1 t
X

X
p

e
g

F
3

3
t

t
X
F

F
p

1
1

1
F
(F)
t
(t)
t F E+i
(t) (F) t (E+i) F
(t) (F)
f f
(t) t (f)
(t) C2 f
f (F)
f (t) C3
(f) C1 f
(t)
F f
f (F) (F)
(F)

Fig. 5. Similarity tree of the questionnaire using new fictitious students (F: Func-
tion, E: Equation, t: technique, i: interpretation, m: modelisation), pointing them
out into squares when they belong to optimal groups regarding the contribution and
typicality (the last one between parenthesis). φ is used to indicate that no fictitious
student belong to the optimal group.
14 m

4 m
6X x

6X x
a pm
g t

e t

gi
3 i

14 i

a pm
g t

e t

gi
3 i

14 i
1 Eg
1p mi

a1 Fp
b Fp
a Fp
3F Fe

2E Fg

1 Eg
0X i

a1 Fp
b1 Fp
a Fp
3F Fe

2E Fg
a6 pi
7e pi

11 pi
1 et
b1 pi

1 m
6E i
7e pi

1p pi

11 pi
5F et
p

p
b1 t
b t

et

10 t

15 t

b1 t
1 t

4 t
b1 t

15 t
p
6E

2E

E
e

17

2E

E
e

F
e
e
3

17

7
t

t
X
4F

9F

8X

9F

F
b6
8

Arbre cohésitif : C:\Documents and Settings\PILAR\Mis documentos\LbroASI\datos-2mod-libro2.csv Arbre cohésitif : C:\Documents and Settings\PILAR\Mis documentos\LbroASI\datos-2mod-libro1.csv

Fig. 6. Cohesion tree under entropic version without (left) and with (right) fictitious
students.
 Fictitious pupils and implicative analysis 335

We display the presence of fictitious students in the optimal groups of

contribution and typicality in Fig. 7.

4 m

X x
a1 pm
g t

e t

i
3 i

4 i

16 Eg
pg
10 mi

b1 Fp
a1 Fp

a1 Fp
3F Fe

2E Fg
a6 pi

11 pi
5F et
7e pi

1p pi

b1 t
12 t

4F t
et

15 t
E
E

E
p

F
e

7
t

t
8X

9F
b6

b1
F
(F) F
(F) t
(t)
E+i
(E+i) f

F+t E
(F+t) F (E+F+i)
f (F)
f (E+i)
F+t
(f )

Fig. 7. Cohesion tree of the questionnaire using new fictitious students (F: Func-
tion, E: equation, t: technique, i: interpretation, m: modelisation). We point them
out into a square when they belong to optimal groups regarding the contribution
and typicality (the last one between parenthesis). φ is used to indicate no fictitious
student belong to the optimal group. Items are labeled as shown in Table 2.

We find again the information previously acquired through the similarity

tree and the cohesion tree under the classical theory: relation between parts ‘a’
and ‘b’ of double questions but, at least, confirms and strengthen the role of
fictitious students as elements helping to display the features of the structures
underlying the variables according to the sample.

4.3 Implicative Graphs

We show the comparative implicative graphs in order to finish the study. En-
tropic formulation is more strict than the classical one regarding the formation
of rules. We switched the originality threshold in CHIC to 0.90 (see Fig. 8).
We observe that both graphs keep a low number of implications at a thresh-
old of 0.99 (the one used under the classical version). The comparison with
the classical setting, in which variations in the implication significances and
336 Pilar Orús and Pablo Gregori

b14Fpi b13Fpi
8Xpmi 5Fet b6Epi b17Egx a17Fgt b13Fpi b14Fpi

11Fet b17Egx 5Fet a14Fet b6Epi a17Fgt a13Fpm

1pt 2Eet 16Xpgi a13Fpm a14Fet

2Eet a6Epi 16Xpgi

Fig. 8. Implicative graphs under entropic version without (left, threshold=0.95) and
with (right, threshold=0.90) fictitious students. Arrows are drawn normal (90%),
barred (95%) and doubled barred (99%) after the intensity the implications exceed.

an arguable choice of the originality threshold (we need to set it up to 0.90

if we want to keep a similar number of implications) dissuades us from using
this analysis.

5 Conclusions
The description of the method involving fictitious students in the different
analyses performed with software CHIC, have permitted us to point out and
show pieces of information which are complementary to the ones resulting
from the classical descriptive statistics and to SIA, that we summarise as
follows:
1. We have ascertained, rather generally, that the introduction of 5 fictitious
students in the initial data matrix has not basically altered the struc-
tures that the different SIA analyses resulted in, just slight modifications
showing a logical but low sensibility in front of a low number of added indi-
viduals (5 over 690 in our case). Then we warn the potential users to check
the size of changes in global analyses before proceeding to the search of
conclusions. This stability in SIA results legitimate, up to our knowledge,
the use the extended matrix in order to interpret results of the original
data through the fictitious students. In that sense, the slight variations
between the analyses mean a positive feedback to the used methodology.
2. Fictitious students, playing the role of students belonging to the opti-
mal group of students regarding either the contribution to the formation
of classes or the typicality within each class, help to explain features of
significant classes issued from the classification and cohesion analysis.
In the first part, using the classical theory of implication, version 3.7
of software CHIC and the a priori matrix made of fictitious students
Calculus, Algebra, exercise, problem, graphic, we have been able to see
that:
 Fictitious pupils and implicative analysis 337

• In the similarity tree, fictitious students representing the type of knowl-

edge (Algebra and Calculus) characterise significant classes.
• Features problem and Algebra do appear together and they characterise
several classes of variables involving items P17a, P17b, P16, P15, P13a,
P13b, whereas features Calculus and exercise also get together for
characterising items P4, P5, P14b and P14a.
• Fictitious student Algebra characterises the implication of P13a, P13b,
P14a, P14b (items included in the same significant class) in the classi-
fication and cohesion analyses. Note that in both cases the implication
involves part ‘a’ and ‘b’ of the same question.
• Similarity and cohesion analyses, with the help of fictitious students,
seem to provide us with right and additional information, mainly about
the pack of items P4, P5, P13a, P13b, P14a, P14b, P17a, P17b, P16,
P15, in the questionnaire. They have proved to be useful in the mining
of our sample.
In the figures displayed throughout the chapter, at the formation of each
class, one can see the codes of items forming the class and the fictitious
students belonging to the optimal groups. As a matter of fact, the high
correlation between what the intuition would tell from the codification,
and what our proposed methodology actually yields, means a support to
the validity of this procedure within SIA, which was motivated by similar
procedures conducted in other multivariate analysis techniques.
in which one can see the codification of items under the criteria used for
the definition of fictitious students, and the presence of fictitious students
in the optimal groups, show a high correlation between what the intuition
would tell (codification) and the results of the positioning of fictitious
students.
3. The conducted analyses have highlighted a certain insufficiency in the
ability of explaining our a priori analysis (choice of features of items), and
in the same way they have given hints for the choice of new convenient
features (fictitious students): type of contents Equation and Function, and
types of task technique, interpretation and modelisation.
4. We conducted a second set of analyses of our data, using the entropic
implication, establishing comparisons with the same analyses conducted
over the extended matrix, remarking that:
• As in the previous case, no significant variations have been produced
in the structure of data, allowing us to use again the interpretation of
fictitious students.
• This second set of fictitious students are not exactly a separation be-
tween type of knowledge and type of task, unlike the previous case.
• They tend to appear alone in the optimal groups, being Function and
technique more present.
• Feature Function emerges related to classes formed by the two parts
of questions 13, 14, 17 (stressing the order b → a) and question 16,
338 Pilar Orús and Pablo Gregori

always significant. Feature technique characterises items P1, P2, P3,

P4, P5, P7 and P10, and the different relations involving them.
• However, also features Equation and interpretation do appear related
to both parts of question 6, stressing the order P*b → P*a.
• The option Typicality, in the cohesion analysis, shows the combination
of these features with other classes.
5. The feasibility study and potentiality of the use of our fictitious students
has been approached by CHIC with the SIA through a mainly descriptive
work. We have intended to highlight the aspect of research tool of this
methodology, and then we have proposed a rather technical presentation
of the use of this tool on the available data.
The necessary small size of the a priori matrix has been determinant in
the final choice of features, although other choices were possible. In spite
of this size, we have found that a considerable amount of information has
been supplied.
In addition, the consideration of two different a priori matrices has con-
tributed to show the changes in SIA results and hence support the idea of
the stability of results of SIA under addition of a small number of individ-
uals in the sample, as well as to appreciate nuances in the interpretation
of results.

References
1. R. Agrawal, T. Imielinsky, and A. Swami. Mining association rules between sets
of items in large databases. In Proc. of the 1993 ACM SIGMOD international
conference on Management of data. ACM Press, 1993.
2. A. Bodin. Modèles sous-jacents à l’analyse implicative et outils complémentaires.
Cahiers du séminaire de didactique de l’IRMAR de Rennes, 1996.
3. G. Brousseau and E. Lacasta. L’analyse statistique des situations didactiques.
In Actes du Colloque Méthodes d’analyses statistiques multidimensionnelles en
Didactique des Mathématiques, ARDM, pages 53–107, 1995.
4. C. Fonseca C, J. Gascón, and P. Orús. Las organizaciones matemáticas en el
paso de secundaria a la universidad. análisis de los resultados de una prueba
de matemáticas a los alumnos de 1o de la uji. In Actas Jornadas de la CV,
Universitat Jaume I, 2002. Societat d’Educació Matemática de la C.V.
5. R. Couturier. Traitement de l’analyse statistique dans chic. In Actes des
Journées sur la Fouille de Données par la Méthode d’Analyse Statistique Im-
plicative, pages 33–50, IUFM de Caen, 2000.
6. R. Couturier and R. Gras. Introduction de variables supplémentaires dans une
hiérarchie de classes et application à chic. In Actes des 7èmes Rencontres de la
Société Francophone de Classification, pages 87–92, Nancy, 1999.
7. J. David, F. Guillet, V. Philippé, and R. Gras. Implicative statistical analysis
applied to clustering of terms taken from a psychological text corpus. In Con-
ference International Symposium Applied Stochastic Models and data Analysis,
AMSDA, Brest, 2005.
 Fictitious pupils and implicative analysis 339

8. C. Fonseca. Discontinuidades matemáticas y didácticas entre la secundaria y la

universidad. PhD thesis, Departamento de Matemática Aplicada, Universidad
de Vigo, 2004.
9. R. Gras. Contribution à l’étude expérimentale et à l’analyse de certaines acquisi-
tions cognitives et de certains objectifs en mathématique. PhD thesis, Université
de Rennes, 1979.
10. R. Gras. Méthodes d’analyses statistiques multidimensionnelles en didactique
des mathématiques. In Actes du Colloque Méthodes d’analyses statistiques multi-
dimensionnelles en Didactique des Mathématiques. ARDM, pages 53–107, 1995.
11. R. Gras and S. Ag Almouloud. A implicaçao estatística usada como ferramenta
em um exemplo de análise de dados multidimensionais. In II Rencontres Inter-
nationales A.S.I. Analyse Statistique Implicative, Sao Paulo, 2003.
12. R. Gras, P. Kuntz, and H. Briand. Les fondements de l’analyse statistique
implicative et quelques prolongements pour la fouille de données. Math. & Sci.
Hum., 154–155:9–29, 2001.
13. R. Gras and A. Larher. L’analyse implicative, une nouvelle méthode d’analyse
des données. Mathématiques. Informatiques et Sciences Humaines, 120, 1992.
14. P. Gregori, P. Orús, T. Bort, I. Pitarch, A. Pérez, J. Gual J, J. García, and
G. Villarroya. Institucionalització departamental d’una prova inicial de matemá-
tiques. In IV Jornada de millora educativa i III Jornada d’harmonització euro-
pea de la UJI, Castelló, 2005. Publicacions UJI.
15. A. Larher. Implication statistique et application à l’analyse de démarches de
preuve mathématique. PhD thesis, Université de Rennes, 1991.
16. I.C. Lerman. Classification et analyse ordinale des données. Dunod, 1981.
17. P. Orús, T. Bort, P. Gregori, I. Pitarch, and G. Villarroya. Evaluación inicial de
los conocimientos matemáticos de los alumnos de primero de la uji. In Actas del
I Congreso de la red Estatal de docencia universitaria y III Jornadas de mejora
educativa, pages 646–665, Castelló, 2004. Publicacions UJI.
18. P. Orús and P. Gregori. Des variables supplémentaires et des élèves “fictifs” dans
la fouille de données avec chic. In R. Gras, F. Spagnolo, and J. David, editors,
Troisièmes Rencontres Internationales A.S.I. Analyse Statistique Implicative,
pages 279–293, Palermo, 2005.
19. P. Orús, P. Gregori, and A. Roig. Observación y producción de conocimientos en
didáctica de las matemáticas mediante la estadística exploratoria. In Actas VII
Simposio de Investigación en Educación Matemática, pages 100–105, Granada,
2003. Universidad de Granada y Valladolid.
20. J. Pearl. Probabilistic Reasoning in intelligent systems. Morgan Kaufmann, San
Mateo, CA, 1988.
21. R.M. Goodman RM and P.Smyth. The induction of probabilistic rule set. the
itrule algorithm. In Proc. of the 6th int. conf. on machine learning, pages
129–132, 1989.
22. F. Spagnolo. L’analisi a-priori e l’indice di implicazione statistica di gras. Quad.
Ricerca Didat, 7:111–117, 1997.
340 Pilar Orús and Pablo Gregori

Appendix

A Theoretical Introduction to Implicative Analysis

and some Inequalities Regarding the Increment
of the Intensity under Addition of New Data

We introduce the necessary notation to present a short theoretical result on

the variations of the intensity of quasi-implications when working under the
classical theory and using Poisson law and a new individual is aggregated to
an existing sample.
Let V be a finite set of binary variables or features (that we shall denote
by a, b, . . .) and E a set of n individuals. Individual x ∈ E is said to possess
(or to be an example of) feature a ∈ V , if a(x) = 1 or a(x) is true.
The rule or implication “a → b” is logically valid whenever for each in-
dividual x ∈ E the logical implication “a(x) → b(x)” is true or, equivalently
when the inclusion {x ∈ E : a(x) = 1} ⊂ {x ∈ E : b(x) = 1} holds. An
individual x ∈ E for which the implication a(x) → b(x) is false, is said to be
a counterexample of the implication a → b.
In the real world, samples rarely offer true logical implications. It is easy to
find counterexamples for any imaginable rule. With the purpose of obtaining
relevant information on the relation among variables, the concept of the rigid
logical implication among variables was extended, weakened, to the concept
of quasi-implication or quasi-rule, more present in real situations [10].
In the sequel we shall work with the single pair of variables a and b. Let
A = {x ∈ E : a(x) = 1} and B = {x ∈ E : b(x) = 1} be subsets of E, and
denote n := card(E), na := card(A), nb := card(B) and na∧b := card(A ∩ B).
Let us now define a random process: let A and B be two random subsets
of E of respective sizes na and nb , whose elements are selected completely at
random from E and independently for each subset.
For any given small α such that 0 ≤ α ≤ 1, it is said that the (quasi-)
implication a → b is admissible at a confidence level of 1 − α when P (card
(A ∩ B) ≤ card(A ∩ B)) ≤ α.
The implication a → b will be admissible at a high confidence level (1 −
α) whenever the chances of finding as many counterexemples or less, in the
random process, as the observed ones in the sample are small (α).
There exists several models describing the random process introduce
above, according to different considerations, such as whether the sample and
population coincides, or whether the size of the sample is fixed or the result of
a random process too (see for instance [2]). We have chosen the one for which
the random variable card(A ∩ B) is distributed as a Poisson law of parameter
n n
λ = an b .
For large values of λ (for instance λ > 5) it is commonly admitted the ap-
proximation
√ by the Gaussian distribution (of mean λ and standard deviation
λ). Then, for the implication a → b, the standarised variable
 Fictitious pupils and implicative analysis 341
na nb
card(A ∩ B) − n
Q(a, b) := q
na nb
n

follows approximately the standarised normal distribution, and its empirical

realisation n n
na∧b − an b
q(a, b) := q ,
na nb
n

expresses the gap between the theoretical and observed values assuming in-
dependence between A and B. This value is called implication index in spite
of being an indicator of the non implication, since it measures the size of
counterexamples.
Now, the intensity of implication a → b is denoted by ϕ(a, b) and defined
as

ϕ(a, b) = 1 − P (card(A ∩ B) ≤ card(A ∩ B)) = P (card(A ∩ B) > card(A ∩ B))

Whenever the Gaussian approximation fits, and therefore the use of the
implication index q(a, b), an approximate value of the implication intensity is
Z ∞
1 2
ϕ(a, b) = 1 − P (Q(a, b) ≤ q(a, b)) = √ e−t /2 dt
2π q(a,b)
Let us show how the intensity ϕ(a, b) varies when a new individual x0 is
added to the sample, in the four different cases (see Table 5). When necessary,
we use subindex 1 with values and random variables concerning the original
sample, and subindex 2 to the respective values regarding extended sample.

x0 a b ∆na ∆nb ∆na∧b λ2 ∆λ

na (nb + 1)
(i) 0 0 0 1 0 >0
n+1
na nb
(ii) 0 1 0 0 0 <0
n+1
(na + 1)(nb + 1)
(iii) 1 0 1 1 1 >0
n+1
(na + 1)nb
(iv) 1 1 1 0 0 >0
n+1
Table 5. Values (or their increments) of the quantities involved in the process of
computation of the intensity ϕ(a, b) when a new individual x0 is added to the sample.

Then we want to estimate ∆ϕ(a, b) := ϕ2 (a, b) − ϕ1 (a, b), hence analyse

the values

P (card(A2 ∩ B2 ) > na∧b ) − P (card(A1 ∩ B1 ) > na∧b )

in cases (i), (ii) and (iv), and

342 Pilar Orús and Pablo Gregori

P (card(A2 ∩ B2 ) > na∧b + 1) − P (card(A1 ∩ B1 ) > na∧b )

in case (iii).
For cases (i) and (iv), where λ2 > λ1 , using the Poisson probability func-
tion and inequalities based on the derivative of the power function, we deduce
P (card(A2 ∩ B2 ) > na∧b ) − P (card(A1 ∩ B1 ) > na∧b ) =
   
 ∞ i i  ∞
 X    X  i i 

λ λ λ λ
e−λ2 2 − e−λ1 1  = e−λ1  e−(λ2 −λ1 ) 2
− 1  ≤



i=na∧b +1 i! i!  i=na∧b +1 i! i! 
   
 X ∞ i i 
 ∞ i−1 
λ − λ X (λ 2 − λ 1 )iλ
e−λ1  2 1 −λ1 2

≤e =
 
i! i!

i=na∧b +1  
 i=0

e−λ1 (λ2 − λ1 )P (card(A2 ∩ B2 ) ≥ na∧b )

and
P (card(A2 ∩ B2 ) > na∧b ) − P (card(A1 ∩ B1 ) > na∧b ) =
   
 ∞  i i 
  ∞  i i 

 X λ λ X λ λ
e−λ2 2 − e−λ1 1  = e−λ2  e−(λ1 −λ2 ) 2
− 1  ≥



i=na∧b +1 i! i!  i=na∧b +1 i! i! 
  ∞ 
 ∞ X (λ − λ )iλi−1 
X λ i − λi 

2 1
e−λ2  2 1 −λ1 1

≥e =
 
i! i!

i=na∧b +1  
 i=0

e−λ2 (λ2 − λ1 )P (card(A1 ∩ B1 ) ≥ na∧b )

In summary, the (positive) increment ∆ϕ(a, b) belongs to interval

∆ϕ(a, b) ∈ [e−λ2 (λ2 − λ1 )(1 − F1 (na∧b − 1)), e−λ1 (λ2 − λ1 )(1 − F2 (na∧b − 1))]
where Fi is the Poisson cumulative distribution function of parameter λi for
i = 1, 2. In case (ii), where λ2 < λ1 , inequalities yield
∆ϕ(a, b) ∈ [−e−λ1 (λ2 −λ1 )(1−F2 (na∧b −1)), −e−λ2 (λ2 −λ1 )(1−F1 (na∧b −1))]
getting a negative increment of the intensity. Although new individual means
no counterexample to the rule, cardinal of B rises (but not cardinal of A) and
then the “surprise” diminishes.
Finally, previous inequalities lead to the following estimation for case (iii):
n +1
λ2 a∧b
∆ϕ(a, b) ∈ [e−λ2 (λ2 − λ1 )(1 − F1 (na∧b − 1)) − e−λ2 ,
(na∧b + 1)!
n +1
λ2 a∧b
e−λ1 (λ2 − λ1 )(1 − F2 (na∧b − 1)) − e−λ2 ]
(na∧b + 1)!
 Fictitious pupils and implicative analysis 343

We encourage researchers to work in this direction in order to find general

bounds for the increment of the intensity of implications, and then providing
SIA users with practical information on how many fictitious students could
be added to a sample in order to have intensity results modified by less than
a given allowed value.

B Questionnaire Driven in the Case Study

Here we show the questionnaire used in the study, translated into English,
and its original version in Spanish.

B.1 Questionnaire (Translation from Spanish into English)

P1 You buy a shirt for PTA 4000 with 15% discount. How much should you
pay for the shirt? 
2x + y = 1
P2 Find out the solutions of the system of equations
3x + 2y = 3
2
P3 Represent the graph for the function t(p) = 4p − p
5
P4 Calculate the derivative of the function f (x) =
(3x − 2)2
P5 Calculate the definite integral (where x is the variable of integration and
R3
a is a constant): 2axdx
1
P6a When solving an equation you get to the expression 0 · x = 8, how do
you interpret this result?
P6b And how do you interpret it when you get to the expression 0 · x = 0?
P7 Calculate the least common multiple of 280 and 350.
P8 A firm is getting an income of I(x) = 50x − x2 USD, where x represents
produced units, and it has expenses of C(x) = 38x + 20 USD. How many
units should it produce in order to get benefits?
P9 The functions f (x) = 3x4 + x and g(x) = x3 − 100x2 tend to zero as x
f (x)
tends to zero. Calculate the limit of the quotient as x tends to zero.
g(x)
P10 How would you compare the following job offers to distribute electoral
brochures? (a) You are paid a fixed amount of PTA 50.000 plus PTA 10
per each delivered brochure. (b) You are paid a fixed amount of PTA
30.000 plus PTA 15 per brochure.
P11 Calculate the derivative of the following function with respect to the
variable x: f (x) = 8sx (where s is a real number)
P12√Can we consider√ both x = 4 and x = 36 as the solutions of the equation
3x − 8 = 4 − x? Provide arguments.
P13a The amount C(t) of water springing from a tap (in litres) is expressed
by an affine function with respect to time t (in seconds). If the water
344 Pilar Orús and Pablo Gregori

gathered in the first second is 3 litres, 5 litres in the second one, and 7
litres in the third one, how much water is gathered in a general instant t?
P13b How much water is gathered in one hour?
P14a The sales of a product after t years from its commercial launch, V (t)
(in thousands of units), is expressed by the function V (t) = 30 · e−1.8/t .
Calculate the limit of V (t) as t tends to infinity.
P14b Interpret the previous result in terms of sales of the above referred
product.
P15 Express in an algebraic language the following sentence: “The product of
three consecutive even numbers is 1287”.
P16 The graphic of the function f (x) = (x − 1)(x + 1)(x + 3), in which points
does it cross the x-axis?
P17a Draw the curves x2 + y 2 = 4, y = 2 − x along the same coordinate axis.
P17b Find out, algebraically, the points where they both intersect.

B.2 Questionnaire (Spanish Original Version)

P1 Compras una camisa que marca 4000 ptas. y te hacen un descuento del
15%. Calcula lo que tendrás que pagar por  la camisa.
2x + y = 1
P2 Busca soluciones del sistema de ecuaciones
3x + 2y = 3
P3 Representa gráficamente la función t(p) = 4p − p2
5
P4 Calcula la derivada de la función f (x) =
(3x − 2)2
P5 Calcula la integral definida (donde x es la variable de integración y a es
R3
una constante): 2axdx
1
P6a En la resolución de una ecuación llegas a la expresión 0 · x = 8, ¿cómo
interpretas este resultado?
P6b ¿Y si llegas a la expresión 0 · x = 0?
P7 Calcula el mínimo común múltiplo de 280 y 350.
P8 Una empresa tiene unos ingresos de I(x) = 50x − x2 dólares, donde x
representa las unidades producidas, y unos costes de C(x) = 38x + 20
dólares. ¿Cuántas unidades hay que producir para obtener beneficios?
P9 Las funciones f (x) = 3x4 + x y g(x) = x3 − 100x2 tienden a cero cuando
f (x)
x tiende a cero. Calcula el límite de la función cociente: cuando x
g(x)
tiende a cero.
P10 ¿Cómo compararías las siguientes ofertas de trabajo de repartir propa-
ganda electoral? (a) Te pagan una cantidad fija de 50.000 ptas. más 10
ptas. por cada papeleta depositada en un buzón. (b) Te pagan 30.000 ptas.
fijes más 15 ptas. por papeleta.
P11 Calcula la derivada de la siguiente función respecto de la variable x:
f (x) = 8sx (donde s es un número real)
 Fictitious pupils and implicative analysis 345
√ √
P12 ¿Se pueden considerar como soluciones de la ecuación 3x − 8 = 4 − x
los siguientes valores, x = 4 y x = 36?. Razona la respuesta.
P13a El volumen C(t) de agua que mana de un grifo (en litros) viene dado
por una función afín respecto del tiempo t (en segundos). Si en el primer
segundo el agua recogida es de 3 litros, en el segundo es de 5 litros y en el
tercero es de 7 litros, ¿cuál es el volumen de agua recogida en un instante
cualquiera t?
P13b ¿Cuál es el volumen de agua recogido en una hora?
P14a La cantidad de miles de unidades vendidas de un producto, V (t), de-
spués de transcurridos t años de su lanzamiento comercial, viene dada por
la función V (t) = 30 · e−1.8/t . Calcula el límite de V(t) cuando t tiende a
infinito.
P14b Interpreta el resultado anterior en términos de ventas del producto en
cuestión.
P15 Expresa en lenguaje algebraico el enunciado siguiente: “El producto de
tres números impares consecutivos es igual a 1287”
P16 La gráfica de f (x) = (x − 1)(x + 1)(x + 3), ¿en qué puntos corta al eje
de las x?
P17a Dibuja las curvas x2 + y 2 = 4, y = 2 − x sobre los mismos ejes de
coordenadas.
P17b Encuentra, de manera algebraica, los puntos donde se cortan.
Identifying didactic and sociocultural obstacles
to conceptualization through Statistical
Implicative Analysis

Nadja Maria Acioly-Régnier1 and Jean-Claude Régnier2

1
EA 3729 University of Lyon — France
[email protected]
2
University of Lyon — France
[email protected]

Summary. To understand culture’s relationship to cognition, this field has studied

children or adults with little schooling and often alien to well-educated Western cul-
ture. Traditionally centered on extra-curricular knowledge, school-based variables
must be considered: written culture and teaching/learning strategies can generate
obstacles to conceptualization. Subjects are adults who studied at least three years
at university: some are professionals. Data was from short clinical-style interviews as
well as a questionnaire based survey taken from an observational sample. To find reg-
ularities linked to conceptual strength, S.I.A. determined implicative rules between
responses and pre-ordered structures. Results suggested representations linked to
specific conceptual aspects constitute didactical and/or socio-cultural obstacles.

Key words: culture and cognition, conceptualization, obstacles, scientific concepts,

prototypical figures.

1 Introduction
In 1744, Tatanga Mani, in his autobiography, “Indian Stoney” commented on
his education “Oh yes! I went to the white man’s school. I learned how to
read his schoolbooks, newspapers and the Bible. But I discovered in time that
this was not enough. Civilized people depend far too much on the printed
page. I turned to the book of the Great Spirit which is present in the whole
of creation. You can read most of this book by studying nature. You know, if
you take all your books and spread them out them under the sun and leave
them for some time, the rain, snow and the insects will do their work, and not
much will remain. But the Great Spirit gave you and I the chance to study
at the university of Nature: the forests, the rivers, the mountains, and the
animals to whom we belong” [20, Mc Luhan, 1971 p.110 in Dasen 2001].
This piece of research is situated within the field that deals with the re-
lationship between culture and cognition [1, 26]. The subject matter takes its
N.M. Acioly-Régnier and J.-C. Régnier: Identifying didactic and sociocultural obstacles
to conceptualization through Statistical Implicative Analysis, Studies in Computational
Intelligence (SCI) 127, 347–379 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
348 Acioly-Régnier and Régnier

source in an academic teaching and learning situation, a course in the didac-

tics of psychology given by Nadja Acioly-Régnier. In other words, the aim is
to reach a better understanding of the clash between cultural rules and ra-
tionality, in particular to grasp more clearly the psychological status of the
procedures and the concepts at work among social actors in various situa-
tions of work or study; here among students of psychology. Some ten years
ago in Brazil, the point of departure was Nadja Acioly-Régnier’s psychology
course for Masters level students and Doctorate seminars at the UFPE3 . The
following situation problem was put forward to the students, based on the
instruction: “Draw the moon as you see it when it is not full.” These students,
being used to beginning their lectures with situation problems concerning the
development of concepts in psychology, did not seem destabilized by this re-
quest. The fact is that the spontaneously produced drawings caused sharp
disagreements between them without reference to the relevant scientific argu-
ments. From the socio-constructivist point of view, the operational concept
of socio-cognitive conflict was at work [9] which in addition Acioly-Régnier
sought to introduce into her courses. Thereafter, this situation was reproduced
in different contexts both geographically and in terms of the public concerned,
to which we will return further in this paper. Strikingly, the answers the stu-
dents provided evoked characteristics similar to those which various studies
had highlighted concerning the performance of illiterate subjects (or of a low
level of schooling) marked by cultural characteristics of group membership, to
the detriment of scientific conceptualization. Thus Luria [21] observed among
illiterate farmers of Uzbekistan and of Kirghizie, a dominant tendency to solve
logical tasks with procedures of argumentation and deduction from their im-
mediate practical experience, when he studied the relationship between the
intellectual abilities of adults and their cultural context.
Scribner [27] observed a propensity in most non-schooled subjects with
little or no education to resort to what she called empirical explanation, as
opposed to theoretical explanation. That is, solving problems relating to syl-
logisms, by resorting to non-relevant or unsuitable reasons taken from out-
side the field corresponding to the problem they had to solve and that these
reasons were subject to strong sub-cultural constraints. The data on which
Scribner’s study is based, amongst other things, relates to subjects of Central
Asia, Western Africa, Mexico and the USA. While seeking to measure the
operational power of the mathematical skills developed out of school in the
sub-culture within which the studied adult subjects worked, Acioly-Régnier
had already made some observations in Brazil. Firstly among sellers of a lot-
tery game called jogo de bicho [1], and secondly, among the Pernambuco sugar
cane workers [2, 3]. Compared with well-educated subjects, illiterate subjects
showed a stronger propensity to avoid confrontation with mathematical prob-
lems, with which they were unfamiliar.

3
Universidade Federal do Pernambuco Recife (Brazil), Town situated in the trop-
ical zone of the southerly hemisphere. (8.03S, 34.54W)
 Identifying didactic and sociocultural obstacles to conceptualization 349

Historically the development of the field indicated by culture and cognition

has essentially relied on work carried out on children and teenagers as sub-
jects [7] or adults with little or no education [16], relating to cultural spheres
often characterized by their differences with the Western dominant culture of
reference, that of the well-read. In this direction, Vergnaud [29–32] noticed
that the study of the relationship between the cognitive and the social had
only just begun. He observed that a great deal of cognitive knowledge which
interests us is cognitive-social (knowledge of children is knowledge with indi-
vidual and social time) and that the process of construction and appropriation
is itself deeply social. Work on social interaction does not constitute a contra-
diction from the constructivist point of view according to which the subject
builds or rebuilds their knowledge. However they make it possible to better
specify the conditions under which this construction is achieved. Vergnaud
adds: “I prefer, for my part, to speak of the process of appropriation of knowl-
edge by the subject, because the knowledge of which I study the training is
socially marked and independent of the subject. A child does not build a sci-
entific discipline, but it does not learn it either without an effort to “rebuild”
it, at least partially”.
The originality of the approach adopted here is that we were interested
in well-educated adult subjects, most of whom are students for at least three
academic years and some of who are already professionals in training or teach-
ing. With regard to the field data, this was collected through short clinical
style interviews as well as a questionnaire-based survey based on an obser-
vation sample obtained by empirical methods. The data resulting from the
questionnaire (appendix 1) are treated in this article by Implicative Statis-
tical Analysis. The aim of data treatment is to start from the concepts of
obstacles to training and conceptual development, to describe and analyze
the various representations of the moon and their effects, as expressed by
Brazilian students in Brazil, by French students of kanaka or caldoche origin
in New Caledonia, by year 6 pupils (9–10 year olds) from Noumea as well
as students in metropolitan France. The existence of different representations
and various capacities to treat a situation is a significant indicator to put
forth hypotheses characterizing the nature of the obstacles which underlie the
representations. From this information, we seek to enhance our understanding
about the nature of the obstacles generated by specific contexts of learning
likely to facilitate or block conceptual development.

2 Cognitive development, learning and obstacles

Our concern in this paper is not to understand how competences are acquired,
but rather to understand how at a given time they are organized by a level
of conceptualization, a notion that underlies the idea of cognitive develop-
ment. Without calling into question Piaget’s model, our approach requires
a theoretical framework that integrates the idea of life long adult cognitive
350 Acioly-Régnier and Régnier

development. We thus conserve the reference to Piaget’s ontogenetic model

whenever we need to consider the interactions between task content and per-
formance. In such cases, Piaget’s model of cognitive development remains a
valuable theoretical reference. Indeed, it can be applied in the adult situa-
tions observed here, if we restrain the sense of this model to one which rep-
resents a partial competence acquisition, rather than full acquisition between
the phases of cognitive development. This reference to cognitive development
underlined also that conceptualization is already an internal construct de-
pending upon processes proper to the individual that are revealed in a social
context, inducing restrictions and limits as much as facilities for the process
of conceptualization.

2.1 Representations, concepts and conceptual development

Vergnaud’s theoretical proposals [32,33] contribute to a better analysis of the

logical bases of representations and the concepts involved, within the frame-
work of the theory of conceptual fields. We consider that the relationship
between reality, signifiers, signification and the concept itself are well sum-
marized by Vergnaud: “Representation is not restricted to a symbolic system
reflects the material world, where signifiers directly represent material objects.
In fact, signifiers (symbols or signs), represent signification which is itself of a
cognitive and psychological nature”. For Vergnaud [30], three levels of entities
should be considered where representation is concerned: signifiers, signification
and reference. The signifiers level consists of various symbolic systems which
are organized differently. The level of signification is central to his theory of
representation, in the sense that it is on this level that invariables are recog-
nized, inferences drawn, actions generated and predictions made: this level
is essentially cognitive. The reference is the real world, as it appears to the
subject’s experience. The subject acts in and on this environment according
to its conscious or unconscious representations. There are three corresponding
problems which must be discussed when speaking about representation: the
relationships between signifiers and signification; between signification and
reference, and that between the various symbolic systems. We hypothesise
that these symbolic systems are arbitrary, that they have no direct relation-
ship with the real world, and already represent a cognitive construct, i.e. that
of signification. Consequently, cultural differences may be better understood
when one considers the problems of the relationships between signifiers and
signification, as well as those relating to the various symbolic systems. Indeed,
the importance given to linguistic meaning or to symbolic systems differs from
one culture to another. Omitting the distinction between signification and sig-
nifiers can lead the researcher to take the symbols and the operations involved
as the essential part of knowledge and cognitive activity, whereas this knowl-
edge and activity are mainly on a conceptual level.
Vergnaud [31] affirms that one cannot have a psychology of complex cog-
nitive activities without knowing what is a concept as a notion, integrated
 Identifying didactic and sociocultural obstacles to conceptualization 351

into that of representation. He defines the concept in the following threefold

way: C= (R, I, S)
R (reference): all situations which give meaning to the concept.
I (signification): all operational invariables on which the effectiveness of
signifiers rests (concept-in-act and theorem-in-act).
S (signifiers): the whole set of signifiers — linguistic and non-linguistic
forms — which makes it possible to represent symbolically the concept, its
properties, the situations and the procedures of treatment. However the sym-
bol is only the “directly visible part of the conceptual iceberg”: the sym-
bolic system is only the directly communicable part of the field of knowledge
which it represents. Lexicon and syntax would be nothing without seman-
tics and pragmatic activities which produce them, i.e. without the practical
and conceptual subjective activity in the real world. From this point of view,
Bachelard (1938/1996) draws our attention to the fact that: “at the same
time, behind the same word, lie so different concepts! What misleads us is
that the same word both designates and explains. Designation is one thing;
explanation is another”.
Concepts are constructs and conceptualization takes place progressively
throughout discovery and the mastery of several kinds of concepts and theo-
rems, most of which remain largely implicit. It is this implicit character which
leads Vergnaud to introduce the notions of concept-in-act and theorem-in-
act [34] and to add: “But an implicit concept is not completely a concept.
It is thus a fundamental theoretical problem to analyze linguistic and non-
linguistic meaning which gives the concept its public character, and which
makes it possible to discuss its definition, its properties, and the truth of the
proposals into which it fits” [33].
A notion which seems to be pertinent to the research presented here is that
of conceptual field . This is a whole, greater than the sum of the situations,
whose mastery requires a specific system of concepts, procedures and closely
connected symbolic notations. It is well known that learning situations do not
always emphasize the same aspects of a concept. This can result in various
levels and various forms of the conceptualization of reality. Such conceptual-
ization can thus appear to be local, which makes it impossible to establish
a relationship between all the elements of a situation, or to recognize these
elements or entities in other situations.
In the second component, I, the concept of signifiers intervenes, the invari-
able construct linked to the behaviour of the subject. From a methodological
point of view, Vergnaud specifies that two levels are to be carefully distin-
guished here:
- the surface level which comprises rules of action and expectations, which
are fairly easily put into words by the subjects. Consequently, this level re-
mains more readily accessible to the observer.
- an in-depth level, containing operational invariables and a system of op-
erations based on these invariables, and which makes it possible to generate
rules of action and expectations. Here, verbalisation by the subject is much
352 Acioly-Régnier and Régnier

more difficult. Consequently, for the observer, this level remains rather inac-
cessible.
In conclusion, in a given situation, the possibility of locating the various
representations which the subjects use to solve a problem can give us im-
portant information to understand their cognitive processing and their level
of conceptualization. In particular, they can show us the obstacles which pre-
vent certain subjects from passing from one type of representation to another,
which may be better adapted and more effective.

2.2 Obstacles to learning and conceptual development

It is now recognised that the production of an erroneous explanation is not

solely due to ignorance, uncertainty, chance or tiredness as the empiricist or
behaviourist theories of learning seem to recommend. These do not recognise
the concept of representation as a relevant and operational concept. Indeed,
it is the well-known works of Piaget and Bachelard which first showed how
errors should be considered as a step towards the acquisition of knowledge, or
as the result of previous knowledge adapted to fit particular circumstances,
but which is presented in the form of false information or simply does not fit
with other circumstances. Thus, in contrast to the empiricist or behaviourist
theories of learning, errors are neither sporadic nor unpredictable, but corre-
spond to representations, and should be thus considered as constructs that
constitute true obstacles which must be overcome.
Within the framework of the didactics of mathematics, Brousseau [5] speci-
fied precisely the notion of obstacle by observing manifestations through errors
bound by a common cause, through ways to conceive a characteristic and co-
herent design, even if it is incorrect, and finally through previous knowledge
which functions within a specific field of action. So overcoming an obstacle
requires an effort comparable to that of learning. Knowledge as an obstacle
is the fruit of the interaction of the individual with his environment, and,
more precisely, with a situation which makes this knowledge relevant. The
variability of human beings, knowledge and the context of learning, leads in-
evitably to the construction of erroneous designs which if locally true, are non
generalizable. However, it should be observed that these designs are guided
by conditions of interaction (individual, medium, knowledge) which function
in a certain way. The identification of these conditions allows the use of di-
dactic objectives. However, these conceptions guided by the parameters of
the interaction (i.e. individual, environment, knowledge) can be modified if
these parameters are well identified and used with didactic aims. It remains
that the obstacles have different origins and content. For our purposes, the
main sources of these obstacles to learning and conceptual development can
be summarized as follows:
• Obstacles of ontogenetic origin linked to the limitation acquisition potential
of the subject during a specific period of cognitive development. Piaget’s
 Identifying didactic and sociocultural obstacles to conceptualization 353

theory offers many examples of the obstacles related to the period of de-
velopment of the subject.
• Obstacles of epistemological origin. These are the obstacles identified
throughout the history of the conceptual development of a discipline. Some
learners’ difficulties may be obstacles which arise from this history. These
obstacles are met for example in the history of science, elements of which
may be observed in the spontaneous models of pupils in learning situa-
tions. An epistemological obstacle is then constitutive of incomplete learn-
ing. This shows that it is not the single result of a chance error which it
is sufficient to correct, or ignorance which can be remedied, or of another
unspecified incapacity. It can result from cultural, social and economic
conditions, but these causes are brought up to date in designs which resist
even when the causes disappear.
• Obstacles which have an educational origin. Bachelard [4] speaks about
the teaching obstacle and Brousseau [5] introduced the concept of an ob-
stacle of didactic origin. As far as we are concerned, we consider that the
conceptualization of reality as well as the representations are constructed
through various types of learning which emphasize certain specific aspects
of reality, and which are themselves related to the particular culture in
which the learning takes place. These obstacles simultaneously come from
different sources: teaching, didactic and socio-cultural and depend upon a
model of teaching and learning.

2.3 The moon phases: an object of study for astronomy

The phases of the moon correspond to apparent changes of this satellite of the
Earth. These changes depend on the respective positions of the Earth and the
Moon in relation to the to the sun. Depending on the time of the month and
year, the sun’s light will shine on one or other part of the moon. When the
moon is between the sun and the Earth, the part lit by the sun is invisible,
giving the new moon. When the first zone becomes visible from the Earth, the
first crescent is visible. When the moon reaches its first quarter, a half-moon
is visible. When the moon is opposite to the sun in relation to the Earth, a
full circle, the full moon may be seen. When the moon reaches three quarters
of its cycle, it is the last quarter, one sees the other half lit by the sun. This
part continues to decrease, the last quarter; then the new moon returns. The
cycle of the phases of the moon, called the lunar month, lasts twenty-nine
and a half days. During the complete rotation of the Earth, the moon also
rotates on its own axis. This brief description (Fig. 1), shows the high degree
of conceptual complexity implied in this process of celestial and astronomical
mechanics. It requires taking into account the moon, the Earth and the sun,
but especially the relationship between three elements, which are in motion.
The moon turns on itself and around the Earth, which turns around the sun
and on itself. The position of the observer on the Earth as well as the time of
day should also be taken into account.
354 Acioly-Régnier and Régnier

Fig. 1. representation of the moon phases

2.4 The moon and its various socio-cultural representations

Throughout history, human beings have always wondered about the moon [15].
We cannot approach in detail all the questions raised here. A number of works
have treated this question, notably, The Moon, myth and image by Jules
Cashford [8]. One of the dominant representations of the moon appears as the
image of a lunar boat crossing the night sky. Figures (2, 3, and 4) show ancient
pictorial representations while the following figures (5, 6, 7 and 8) show visual
representations of the moon with which subjects may be confronted in their
actual daily life. Drawings (6, 7) are taken from comic strips produced in the
southern hemisphere respectively by Mauricio de Sousa (Fig. 6), in Brazil and
by Bernard Berger in New Caledonia (Fig. 7).
It is also necessary to add all the scientific photographs produced by satel-
lites or even by man’s direct visits, on several occasions, to the moon itself.
There are also those produced synthetically by computers.
We can find artistic representations of the moon’s phases such as figure 8

2.5 Phases of the moon: an object of everyday learning

The moon in all its apparent phases constitutes an object of the perceptive
experience of everyday life in early childhood. The observation of this object
 Identifying didactic and sociocultural obstacles to conceptualization 355

Fig. 2. Entry of Venus’ Sanctuary, Paphos (Harding 2001 p.92)

Fig. 3. Assyrian winged moon (Harding 2001 p.96)

Fig. 4. Lunar tree surrounded by lattice and torches (Harding 2001 p.93)

Fig. 5. Black/white creation Editions M.D.

Fig. 6. Chico Bento (number 163, 1993) Mauricio de Sousa Editora GLOBO São
Paulo Brazil
356 Acioly-Régnier and Régnier

Fig. 7. Small/large Boat. The Bush in Madness n◦ 10 (1996/2002) Bernard Berger

Edition Noumea New Caledonia

Fig. 8. Lunation 1990 Photographs by Rimma Gerlovina & Valery Gerlovin

constitutes then a first experience in the Bachelardian sense. The mental rep-
resentations that the subjects construct in this way, can constitute obstacles
of epistemological origin with which they will be confronted when it is a ques-
tion of understanding the phenomenon of the phases of the moon. From a
Vygotskian understanding, this observation leads to the spontaneous forma-
tion of daily concepts relating to the phenomenon of the apparent movement
of the moon and of its phases, and its link with the apparent movement of
the sun, with a weak use of language. These concepts are isolated from each
other and develop apart from any given system. They are temporally or lo-
cally relevant, and may also lead to generalizations which can be abusive.
Their conceptual weakness is manifested by an incapacity for abstraction or
an inaptitude for intentional use. What is characteristic is their incorrect use.
Frequently saturated by the rich personal experience of the subject, and as
such, socio-cultural obstacles to conceptual development become apparent.
Data resulting from discussions with eight adult illiterate subjects of rural
areas of Brazil illustrates our research. These subjects are characterized as all
having little contact with a written culture. Questioned on “how they see the
moon when it is not full ”, they provide answers giving clues to the nature of the
obstacles indicated in this article. Thus Maria, 60 years old, a cleaning lady,
from the city of the Sertão, Nordeste of Brazil, draws the moon B (Appendix
1) and explains that “it is like Lampião’s hat” (a famous character known
in that city). Nen, 40 years old, with the drawing of the moon C (Appendix
1), explains why “it is like a smile”, and Neta, 35 years, that “the moon is
presented in the form of a hammock”. This complementary data consolidate
 Identifying didactic and sociocultural obstacles to conceptualization 357

our interpretation of the local aspect of knowledge of the illiterate subjects,

but considered further also seem to reveal the effects of what is learned at
school, as obstacles to processes of conceptualization.

2.6 Phases of the moon: an object of learning at school

Concerning scientific concepts, Vygotski postulates that they arise from indi-
rect contact with the object and can be acquired only by a continuous process
from general experience to the private individual. They are formed under the
intentional action of the school, beginning with a teacher’s explanation, who
exposes a scientific formulation of the concept. Their weakness lies on the one
hand in his verbalism, principal source of the shifts which generate obstacles to
conceptual development; and on the other hand, insufficient links to concrete
experience and knowledge. However, analysis of our research data corroborates
with Piaget’s idea (1969) where verbalism of the image also exists.
Through a survey carried out over the last five years among primary school
teachers still in training and those in secondary schools in initial or profes-
sional training at the University Teacher Training Institute (IUFM) of Lyon,
we collected data which enabled us to identify three main categories of ap-
proaches to teaching the question of the phases of the moon in their courses
in France. The first was described as a way of telling a fable or tale to ap-
proach the concept of the phases of the moon for young children. The second
was based on more technical elements calling upon mnemonic techniques that
they themselves had acquired during their own education and more related to
learning at secondary school. Finally the third target group was also primary
school teachers and was based on a model implying a scientific approach to
observation similar to that of astronomers. In an article on Toussaint (1999)
we found an extremely relevant and adequate description similar to that which
we observed among teachers’ replies, related to the first two steps

First approach: “the moon tells lies. . . !”

The concept is introduced through a fable or tale. The following instruction is

given to the pupils, and may be thus summarized: “Don’t forget that the moon
tells lies: when it looks like the letter C, its not waxing, its really waning;
and when it looks like the letter D, its waxing”. (In French, or other Latin
languages, C represents the first letter of the verb croître — to inCrease.
The letter D, décroître, means Decrease). This aim of this approach is clearly
to help recall and to understand the phases of the moon. The emphasis is
on the signifiers the first letters of the verbs, which enables recognition and
verbalisation of the different phases. It must be noted that this approach does
not in any way reflect the conceptualization of the astronomical movement
of the moon. The accent is only on meaning, which makes it possible for
the subjects to recognize a position of the moon without taking into account
the dynamic relationship of the concept involved. In addition, the concept
358 Acioly-Régnier and Régnier

associated only with only one situation is not restricted here by one meaning to
give an account of all the representations which the subjects by will have built
elsewhere through early experience by looking at the moon in their childhood.
This story-telling approach does not give optimum conditions for learning
subjects to go beyond the everyday concepts relating to the phases of the
moon, to arrive at scientific concepts with the meaning given by Vygotski [35].
In addition the images given by the forms D and C constitute prototypical
representations of phases of the moon within Rosch’s meaning [25].

Second approach: “the moon is no longer a liar. . . !”

Here the question of the phases of the moon is introduced with more technical
concepts. Toussaint [28] gives an account of it by calling upon its characteris-
tics experienced by a schoolboy.
“Later, I came across another rule which said that the moon (it was less
funny!) was no longer telling lies: by carefully drawing the diameter which
goes from one end to the other, one can write in tiny characters a P with the
first quarter and a D with the last”. We see that in this case a geometrical
concept appears — diameter — which can give the impression of a more
learned approach. But the use of diameter does not bring anything more than
the mapping of a lunar position with one of the two letters P and D, which are
mobilized only as meaning in a way identical to the first step. Surprisingly, the
teachers however, seem to regard this approach as requiring a higher level of
conceptualization. This is suggested by the fact that this approach was never
used with primary school pupils.

Third step: towards a scientific model of the phases of the moon. . .

This less frequent third approach called upon systematic observations and the
use of a model with the pupils. We collected results from a teacher training
course on the didactics of physics. In this step, they were asked to identify the
representations of the phases of the moon, then to carry out systematic ob-
servations of the moon with which the representations are confronted. Finally
the trainee teachers are confronted with a model in the shape of a model of
the system Earth-moon-sun which reproduces the movements of these bodies.
Starting from this mechanical device, they are confronted with problems of
the type: in its various phases how does the moon appear from the Earth?
It appears that this model, in spite of its concrete material characteristics,
allows only a limited understanding for trainee teachers, insofar as when they
approach the phases of the moon they do so infrequently or not at all in their
teaching practice. The main argument they use relates to the high cost of such
device in a teaching-learning situation. However it is clear that this step, as
“costly” as it is, requires an active process of conceptualization.
 Identifying didactic and sociocultural obstacles to conceptualization 359

3 Identifying the conceptualization levels

and the associated obstacles
As we said in the introduction to this article, the central question of this
study is to identify the conceptual levels and the resulting obstacles that
are associated with the conceptualization of moon phases. The starting point
of the study was the observation of learning-teaching situations for Masters
and Doctorate psychology students in Recife, Brazil. Thereafter, the reference
field of these problems was extended to other learning-teaching situations in
France and in New Caledonia and involved Bachelor’s and Master’s students
of Educational Sciences, students training to be educational psychologists,
social workers in university training in Educational Sciences, and also teacher
trainees in initial training and even teachers in professional training. In all
the contexts in which Nadja Acioly-Régnier led these teaching sequences, the
teaching situation invariable lay in the injunction to draw the moon as each
one could see it when it is not full. Interestingly, in all the situations, the main
characteristic of the drawings was that they had a closer relationship with the
representations of written culture than by direct observations carried out by
the subjects. As for the arguments put forward by the subjects, their main
characteristic can be summarized as follows: the drawings are more related to
the images we perceive in our socio-cultural environment than what we observe
directly by looking at the sky.

3.1 An initial situation problem focussed on the phases

of the Moon

To begin with, we were struck by the fact that the Brazilian students faced
with the instruction to draw the moon as they saw it from the subequatorial
tropical zone, systematically produced drawings with a typical representa-
tion of the moon observed more frequently from mainland French sky. The
great majority presented the moon in the shape of a “crescent” (Fig. 6) facing
the right and a lower proportion in the shape of a “crescent” facing the left
(Fig. 5). The contradictions of these two categories of response give rise to
a genuine situation of socio-cognitive conflict. However by remaining on this
level of exchange, the nature of the responses built on the basis of first-hand
experience and on the tools provided by the cultural environment and the
written culture did not offer the conditions for a significant rise in the level
of conceptualization. To modify these conditions, the subjects were then in-
vited to directly observe the moon and to compare their perception with their
drawings. The result then seemed to provoke a real cognitive destabilisation
and a desire to understand the situation from a conceptual point of view.
Within the context of psychology teaching, this didactic situation carried out
on the phases of the moon led the students to analyse the school textbooks
to pinpoint the role and the place of iconic representations in scientific learn-
ing. Piaget himself [23, p.110] considered that: “image, film and audiovisual
360 Acioly-Régnier and Régnier

processes of which all pedagogy keeps harping on about today and wants to give
the illusion of being modern, are invaluable auxiliaries as an addition or as
a spiritual crutch, and it is obvious that this is a progress compared to purely
verbal teaching. But there is verbalism of the image just as there is verbalism
of the word”.
This activity had led the students to verbalise their awareness that the
responses initially provided were determined by school learning which, in
the name of immediate efficiency, favours excessive simplification and reduces
memory based learning of the signifiers without working on the concepts to
which they are attached. This perspective is reinforced, in the cultural en-
vironment, by the graphic representations of the moon that we find in the
media, in comics (Fig. 6 and Fig. 7), in advertisements, etc.
When we transferred the situation in France to the Teacher Training Insti-
tute (Institut Universitaire de Formation des Maîtres) in Lyon, as well as at
the university (University Lyon2), we modified the procedures. The situation
problem given to the students and the trainee teachers was not based on the
request of drawing the moon, but on a story told as follows: “In Brazil, we
asked to students to draw the moon as they saw it in their sky. Because they
drew the moon as a vertical crescent or a slightly bent one, that faced the right
for the most part of them or the left for the others, I asked them to observe
it directly by watching in the sky. I told them to increase the stakes by paying
for a coffee to anyone that will see it as he drew it”. To end the story we
added: “Nobody came to claim the offer ”, and we then addressed the following
question to the French students: “Why was this the case?”
Using a qualitative approach, the analysis of the responses obtained lead
to differentiate five main categories. The subjects were allowed to give several
answers in their arguments and also to change the meaning of the answer by
exchanging their views with their fellow students.
The first category [CAT1] corresponds to responses overly determined by
the socio-cultural dimension. In this category we found the following: “They
did not dare to do so because you are married” or “. . . you are a professor” or
“ . . . your husband is jealous”, and also “Because they watched the moon lying
in a hammock” etc.
The second category [CAT2] corresponds to responses overly determined
by the personal characteristics of the subjects questioned. For example: “They
disliked coffee, or pubs” etc.
The third category [CAT3] corresponds to responses overly determined by
local conditions, contexts, or circumstances. We found for example: “they were
living downtown and could not see the sky so well” or “there was fog” etc.
The fourth category [CAT4] corresponds to answers overly determined by
the exact knowledge of the situation-problem. For example: “I cannot say, I’ve
never been there” or “I went there, and obviously the moon is not like that”
etc.
The fifth category [CAT5] corresponds to responses overly determined by
school knowledge. We found for example: “I believe it has something to do
 Identifying didactic and sociocultural obstacles to conceptualization 361

with shadows, and light” or“. . . with the moon, the sun, and the earth and the
hemispheres” etc.
As said in the introduction, Luria [21], studying the illiterate Uzbek farm-
ers without any written culture, observed similar results. The question is then:
Are the facts here collected just anecdotic, representing particular behav-
iour of subjects confronted with a scholarly situation-problem even though
they are at a university level during their initial training?
Alternatively, do these facts reveal socio-cultural obstacles, impeding the
rise of the level of conceptualization, with pedagogic and/or didactic origins
that could be found in the school institution, or with epistemological origins
that are linked to the development of the concept? In order to explore such
hypotheses further we designed a questionnaire aimed at investigating mental
representations concerning the moon phases and we conducted a survey on
subjects from various geographical and socio-cultural contexts.

3.2 Exploration of the mental representations of the phases

of the moon
The first questionnaire was drawn up taking into account the observations
previously evoked. It comes within the perspective of a structure of data
founded on recognition of static shapes proposed a priori to the subjects and
no longer on the production of a graphic shape supposed to represent the
moon for the subjects. The four shapes A, B, C, D (Fig. 3.2) are formulated
in relation to question Q1, Q2, Q3, and Q6, reproducing the graphic shapes
found in written work in the course of history and which have been able in turn
to be prototypical shapes in various eras and contexts. Question Q4 and Q5
repeat the pedagogical approach described by professors in charge of teaching
this content. Furthermore, the subject who answers is the subject directly
concerned with the instruction in question Q1. In questions Q2 and Q3 the
reference point of a particular subject to which the subject who answers must
put himself or herself in their place, and aims to introduce the variability
of the observer’s position on the surface of the Earth. So various forms of
the questionnaire were submitted in such a way that in question Q3 it was
either the actual geographic location of the subject who answered or it was
geographically opposite in question Q2. The aim of question Q6 is to bring
out the representations that the subjects have with respect to the perception
of the moon from each terrestrial hemisphere and from the equator.

Questionnaire-based survey
This questionnaire-based survey was carried out along with observation ob-
tained through empirical methods located respectively in metropolitan France
in Lyon, in Noumea, New Caledonia and in Recife, in north-east Brazil over a
long period from 2001–2004. The overall sample is composed of 198 subjects.
Note that the individuals of the Ech_IUFM sample were not submitted
that to two questions Q1 and Q2
362 Acioly-Régnier and Régnier

Representations of the moon

Answers Q1; Q2; Q3; Q6

A-Yes A-No B-Yes B-No C-Yes C-No C-Yes C-No
[1Asim] [1Anao] [1Bsim] [1Bnao] [1Csim] [1Cnao] [1Dsim] [1Dnao]

Fig. 9. Shapes of the moon

New Caledonia Metropolitan Recife, in North-

France East Brazil
Adults Children Adults Adults
Students in Master Elementary school Trainee teachers: Curso normal supe-
of educational sci- students: grade 6 in 2nd year in IUFM rior da UFRPE
ences; situation of Noumea de Lyon
training
Adults professional Children Adults in situation Adults in situation
in the educative of professional of professional
system, as teacher training (teachers) training (future
or as responsible of teachers)
teacher’s training
EchNC_Ad EchNC_Enf Ech_IUFM Ech_UFRPE
119 22 28 29
Q1; Q2; Q3; Q4; Q1; Q2; Q3; Q4; Q1; Q2 Q2; Q3; Q4; Q5; Q6
Q5; Q6 Q5; Q6
Table 1. Constitution of the overall sample in function of the contexts

Treatment and analysis implicative statistics: modelling

and description

The first part of the questionnaire consists of 7 questions which provide infor-
mation relating to the subjects such as sex, age and the length of professional
experience. The variable SEX corresponds to the vector variable (MALE;
FEMALE) whose two components are additional binary variables. We also
introduced a “place of residence” variable whose detailed form corresponds to
the membership of a “usual cultural area: Kanak” in New Caledonia. However
we restricted it here to the couple of additional binary variables (HN; HS)
indicating respectively the northern and southern hemispheres. The variable
“Age” was modelled by the couple of additional binary variables (Child; Adult)
The question Q1 concerning the recognition or not of the shapes and ap-
parent positions of the moon is represented by a variable vector with 12 binary
components (1Asim; 1Anao; 1Anr; 1Bsim; 1Bnao; 1Bnr; 1Csim; 1Cnao; 1Cnr;
1Dsim; 1Dnao; 1Dnr). We coded xxnr the absence of answer.
 Identifying didactic and sociocultural obstacles to conceptualization 363

Sample EchNC_Ad EchIUFM EchUFRPE EchNCEnf

SEX Male Female Male Female Male Female Male Female
Number of responses SEX 50 68 4 24 4 25 15 6
Number of responses SEX
and AGE 47 60 4 24 2 25 15 6
AGE in Years Min. 23 27 24 22 18 17 10 10
Max. 49 51 33 45 32 40 12 12
Mean. 38.5 37.6 29 26.4 25 23.0 10.7 10.6
St.dv 5.2 5.6 3.7 5.9 7 6.2 0.68 0.70
Number of responses
SEX and LENGTH 48 66 X X 2 18 X X
Length of professional
experience in Years Min. 2 3 X X 5 0 X X
Max. 27 31 X X 10 20 X X
Mean 14.5 14.3 X X 7.5 2.85 X X
St.dv 5.57 6.54 X X 2.5 4.81 X X
Table 2. Description of sample (Numbers of responses, measures of location and
dispersion)

The questions Q2 and Q3 which urge the subject to be located from an-
other point of view, each one are represented by a variable vector with 15
binary components:

Q2 = (2sim; 2nao; 2nr; 2Asim;2Anao; 2Anr; 2Bsim;

2Bnao;2Bnr; 2Csim; 2Cnao; 2Cnr;2Dsim; 2Dnao; 2Dnr)
(1)
Q3 = (3sim;3nao; 3nr; 3Asim; 3Anao; 3Anr; 3Bsim;
3Bnao; 3Bnr; 3Csim; 3Cnao; 3Cnr; 3Dsim; 3Dnao; 3Dnr)
The questions Q4 and Q5 relate to judgements of effectiveness and adap-
tation of a teaching approach based on the correspondences of the phases of
the moon with letters C, D, p or D. They are both modelled by a binary
variable vector of dimension 4

Q4 = (4ADAP; 4INAD; 4EFFI; 4INEF)

(2)
Q5 = (5ADAP; 5INAD; 5EFFI; 5INEF)
Finally the last question Q6 relates to taking into account the three places:
southern hemisphere, northern hemisphere and equator. It is modelled by a
binary variable vector of dimension 15:

Q6 = (6sim; 6nao; 6nr; 6A_HN; 6A_E; 6A_HS ; 6B_HN; 6B_E; 6B_HS;

6C_HN; 6C_E; 6C_HS; 6D_HN; 6D_E; 6D_HS)
(3)
The frequency distributions of these variables are provided in appendices
(Appendix 2, Appendix 3)
364 Acioly-Régnier and Régnier

The modeling of the questions by variable vectors with binary components

enables us to place ourselves in the context of statistical implicative analysis,
SIA developed by Gras Régis and his collaborators [11, 12, 24] starting from
prospects released by I. C. Lerman [19] and processed by CHIC software
(Couturier). The table of the statistical series is a table of 67 columns made
up of values 0 or 1 describing a realization of the 67 binary variables.
Seven open questions formulated from the interrogative one: Why? give
place to textual answers and are thus modelled by textual variables. At this
stage of our work in this article we do not proceed to a refined computer-
assisted analysis of contents using software such as SPAD_T4 [17] which could
be also supplemented by the SIA approach.

Questions Yes Not Not-Reference Total Yes(%) Not (%) Not-Reference

mark mark (%)
Q1 1A 171 21 6 86.36 10.61 3.03
1B 58 116 24 29.29 58.59 12.12
1C 65 110 23 32.83 55.56 11.62
1D 160 25 13 198 80.81 12.63 6.57
Q2 2 163 34 1 82.32 17.17 0.51
2A 40 120 38 20.20 60.61 19.19
2B 109 48 41 55.05 24.24 20.71
2C 112 40 46 56.57 20.20 23.23
2D 54 107 37 27.27 54.04 18.69
Q3 3 114 51 5 62.18 34.45 3.36
3A 31 102 37 18.49 63.87 17.65
3B 71 61 38 38.66 42.86 18.49
3C 65 67 38 170 38.66 42.86 18.49
3D 30 102 38 18.49 63.03 18.49
Table 3. Frequency Distributions of answers to Q1, Q2, Q3.

Results (Table 3) relating to Q1 (Did you already see the moon like that
in reality?) confirm the use of prototypical memory (within the meaning of
Rosch [25]) in the evocation and the recognition of these lunar forms. Ac-
cording to Eleanor Rosch, among all the levels of abstraction possible, one
is psychologically more accessible than the others: called “basic level”, level
which makes it possible for the individual to obtain the maximum informa-
tion with the minimum of cognitive effort. Compromised between the most
abstract possible level but which offers, in same time, a sufficient number of
concrete attributes. Thus, the figures 1A (86.36%) and 1D (80.81%) “growing
directed vertically”, seem figures prototypic independently of the group studied
(Test of χ2 Appendix 4) with a prevalence of 1A “horns directed towards the
line”. On the other hand the figure 1B (29.29%) “horizontal crescent turned
4
SPAD_T Système Portable d’Analyse des Données Textuelles CISIA-France
 Identifying didactic and sociocultural obstacles to conceptualization 365

downwards” and the figure 1C (32.83%) “horizontal crescent turned upwards”

are quoted little by the subjects. These two variables depend on the variable
“groups” (Appendix 4) with one attraction towards 1B and 1C for EchNC_Ad
and one repulsion for the three other groups.
It may be noted that no significant dependence (with the test of χ2 with
α = 0.05) is detectable between the variables “Sex”, “Age” “Hemispheres” and
the variables 1A, 1B, 1C and 1D.

Q1 (A) Q1 (B) Total

SEX 1Asim 1Anao 1Anr 1Bsim 1Bnao 1Bnr
Male 59 12 2 24 40 9 73
Female 110 9 4 33 75 15 123
169 21 6 57 115 24 196
df. = 2, k = 5.99 χ2 = 3.99 χ2 = 0.88
Q1 (C) Q1 (D) Total
SEX 1Csim 1Cnao 1Cnr 1Dsim 1Dnao 1Dnr
Male 25 39 9 57 12 4 73
Female 39 70 14 101 19 9 123
64 109 23 158 31 13 196
df. = 2, k = 5.99 χ2 = 0.23 χ2 = 1.56
Table 4. bivariates Distributions Q1 and SEX.

The objective of Q2 (“if a subject living in the opposite hemisphere to you,

says to you that he/she had never seen any of these moons in reality, would
you believe him/her?”) and that of Q3 (“If a subject living your hemisphere,
says to you that it had never seen some of these moons in reality, you believe
it?”) were the introduction of a data being able to draw the attention of the
subject to conceptual aspects of the phases of the moon. It could thus oppose
its own answers to those of this virtual subject, either by maintaining those
brought to Q1, or not answering. We find results similar to those which we
pointed for Q1. We find there also initial classification [CATn] (N = 1 to 5)
exposed into 3.1.

Exploration of the trees of similarities and cohesions, and the

implicative graph

Initially, we studied the data built on the 27 principal binary variables (Ap-
pendix 2) starting from the sample of the 198 individuals of which 28 lived
in the northern hemisphere and 170, the southern hemisphere. We gave to
a second publication, the exploitation of the 67 basic variables (Appendix 2,
Appendix 3) on the sample of 170 individuals of the southern hemisphere. In
fact here we studied the 27 variables which relate to the whole of the total
sample.
366 Acioly-Régnier and Régnier

Fig. 10. Tree of similarities

By classification based on a model inspired by Lerman using probabilistic

indices, we obtain a repartition of the 27 binary variables in four main classes
as the tree of similarities shows (Fig. 10).

CLS1(lev22) = {1Asim, 1Dsim, 1Bnao, 1Cnao}

CLS2(lev23) = {1Anao, 1Dnao, 1Bsim, 1Csim, 2nr, 2Asim
2Bnao, 2Cnao, 2Dsim} (4)
CLS3(niv21) = {1Anr, 1Dnr, 1Cnr, 2nao, 2Anr, 2Bnr, 2Cnr, 2Dnr}
CLS4(lev20) = {2sim, 2Anao, 2Dnao, 2Bsim, 2Csim}

This repartition reflects a control of completely coherent answer. Each

class consists of methods which are logically associated
These classes result from the aggregation of binary variables explainable
by the effect of the prototypical figures in the reading of the world by the
subjects. Thus the figures A and D, on the one hand, B and C, on the other
hand are strongly associated for Q1 and Q2.
The analysis of the contributions of the subjects are characterized re-
spectively by the categorial variables “Sex”, “Age”, “Hemisphere” additionally
though the optimal groups do not reveal outstanding effects. At most we ob-
serve a significant influence of the binary variable CHILD in the constitution of
class CLS1 (lev22). That could correspond to the importance of the education
 Identifying didactic and sociocultural obstacles to conceptualization 367

with which they are confronted daily at their age and, in particular, figures
that they meet in their textbooks or even in children’s books.
Let us explore the implicative graph built to leave the 27 binary variables
instanced on the total sample of 198 individuals. With a confidence level of
0.99, we identify 7 implicative chains (Fig. 11). Five comprise only 2 terms.

Fig. 11. Implicative Graph, [Confidence Level 1 − α = 0.99]

(ch1) [1Dsim] ⇒ [1Asim]

(ch2) [1Anao] ⇒ [1Dnao]
(ch3) [1Bsim] ⇒ [1Csim]
(ch4) [1Cnao] ⇒ [1Bnao]
(5)
(ch5) [1Cnr] ⇒ [1Bnr]
(ch6) [2nao] ⇒ [2Dnr] ⇒ [2Anr] ⇒ [2Bnr] ⇒ [2Cnr]
(ch7.1) [2Dnao] ⇒ [2Bsim] ⇒ [2Csim] ⇒ [2Anao] ⇒ [2sim]
(ch7.2) [2Asim] ⇔ [2Cnao] ⇒ [2Bnao] ⇒ [2Dsim] ⇒ [2sim]
Just as we saw we in the tree of similarities, what emerges from these
chains of quasi-implications that the subjects answered (yes, no, failure to
reply) by strongly associating the figures A and D, on the one hand, and the
figures B and C on the other hand, as well within the framework of Q1 as
Q2. The two twin variables (2Asim) and (2Cnao) also illustrate this property
by the strong association of opposition between the positive designation of a
prototypical figure A and that negative of the not-prototypical figure C. We
observe in more than one almost perfect inclusion of {2Asim} in {2Dsim}
insofar as there is one counterexample. This always confirms the dominant
place of the prototypical figures in the mental representations and their role
368 Acioly-Régnier and Régnier

in the construction of the obstacles to the development of conceptualization.

In our case, this relates to the acquisition of a representation of the world —
here the moon — which is closer to that elaborated in learned knowledge.

Chain (Chx)
(CH1) (CH2) (CH3) (CH4) (CH5) (CH6) (CH7.1) (CH7.2)
Questions Q1 Questions Q2
The most typical variable
Child Male Male Child HS Child Female HN
Level of risk incurred by choosing the variable like most typical
0.149 0.117 0.278 0.0952 0.224 0.112 0.194 0.0355
The most contributive variable
Child Male Male Child HS Child Female Male
Level of risk incurred by choosing the variable like most contributive
0.149 0.117 0.278 0.0952 0.224 0.1 0.236 0.149
Table 5. Contributions and typicality of the additional variables

By considering the weakest risk of the level, we observe that group HN

is most typical of the chain (Ch7.2). It corresponds in fact to that made up
by IUFM teacher trainees in Lyon, only located in the northern hemisphere
and characterized by the highest academic level. The semantic characteristic
of this chain is the acceptance of a different perception of the moon according
to the hemispheres and the association of prototypical figures between them
(A and D accepted) and, of the other, those which are not-prototypical (B
and C rejected). Effects resulting from the courses on the didactics of physics
are a possible interpretation of the amount of distance from these subjects.
So now we explore the cohesitive (implicative) tree, it gives six main classes
from binary variables.

C1_lev21(5) = ((1Anao ⇒ 1Dnao) ⇒ (1Bsim ⇔ 1Csim)) ⇒ 2nr

C2_lev20(4) = ((1Anr ⇒ 1Dnr) ⇒ (1Bnr ⇔ 1Cnr))
C3_lev17(4) = ((1Bnao ⇔ 1Cnao) ⇒ (1Dsim ⇒ 1Asim))
(6)
C4_lev9(5) = (((2nao ⇔ 2Anr) ⇔ 2Bnr) ⇒ 2Cnr) ⇔ 2Dnr
C5_lev16(4) = 2Cnao ⇒ ((2Asim ⇔ 2Dsim) ⇒ 2Bnao)
C6_lev14(5) = 2Dnao ⇒ (2Csim ⇒ (2Bsim ⇒ (2Anao ⇒ 2sim)))

In this classification, we find the classes which were formed in the approach
by similarity. The formed R-rules are organized in a structure which confirms
the properties of the prototypic representations and the obstacles that they
induce such as we identified them through the analysis of the implicative
graph.
The R-rules are conceived as an extension of the binary rules (A) ⇒ (b)
with the rules. They are affected to a d°R degree calculated according to the
 Identifying didactic and sociocultural obstacles to conceptualization 369

Fig. 12. Implicative Tree

following definition (Kuntz ASI 2005): a R-rule composed of a binary variable

is degree d°R=0. A R-rule consists of a binary rule with a degree d°R=1. A
R-rule (R0 ) ⇒ (R00 ) admits a degree d°R = d°R’ + d°R” + 1.
By paying attention to the coherence O(C) [13] of a class C representing a
R-rule to be left confrontation enters the order of the binary variables deter-
mined by the occurrences and that determined by the rules within the class.
This coherence is measured starting from the inversions by the probability P
{I> i} where i is the number of inversions observed and I the random variable
“numbers of inversions”
For the class C1_lev21 (5) we obtain the order (1Anao, 1Dnao, 1Bsim,
1Csim, 2nr) within the class while the order resulting from the occurrences
would give (2nr, 1Anao, 1Dnao, 1Bsim, 1Csim). We thus count 4 inversions
between these two permutations. We established [13, p. 44] that the proba-
bility to have 5 or more inversions is 71/120, that is to say approximately
59.16%.
Notice that the three classes C1, C2 and C3 result from binary aggregations
of variables associated with Q1, whereas the three others are associated with
Q2.
When we study the typicality of each class of this division starting from
the additional variables, we obtain the results in Table 6.
The C6 class incorporated at the significant level 14 consists of the vari-
ables of the chain (Ch7.1) and a R-rule of degree 4 represents. The order
in the class leads to the permutation (2Dnao, 2Csim, 2Bsim, 2Anao, 2Sim)
370 Acioly-Régnier and Régnier

Classify
C1_lev21 C2_lev20 C3_lev17 C4_lev9 C5_lev16 C6_lev14
(5) (4) (4) (5) (4) (5)
degree of R-rule
4 3 3 4 3 4
Coherence
71/120 ≈ 20/24 ≈ 20/24 ≈ 91/120 ≈ 20/24 ≈ 115/120 ≈
0.5916 0.8333 0.8333 0.7583 0.8333 0.9583
The most typical variable
MALE HS CHILD CHILD HN FEMALE
Level of the incurred risk
0.278 0.224 0.0334 0.1 0.0355 0.229
The most contributive variable
FEMALE HS CHILD CHILD HN FEMALE
Level of the incurred risk
0.315 0.365 0.0361 0.1 0.0833 0.236

Table 6. Contributions and typicality of the additional variables

while the order resulting from the occurrences gives (2Dnao, 2Bsim, 2Csim,
2Anao, 2Sim). These two permutations reveal an inversion. From there co-
herence O(C6) ≈ 0.9583. The FEMALE characteristic contributes more and
remains most typical of the C6 class like chain (Ch7.1). At this level of infor-
mation, we do not have a relevant interpretation of this class in relation to
the contribution and the typicality of the female group.
The C5 class incorporated on level 16 corresponds to a R-rule of degree
3 and is associated with the under-chain (Ch7.2a) [2Asim] ⇔ [2Cnao] ⇒
[2Bnao] ⇒ [2Dsim]. Its composition still reflects the association of the pro-
totypical figures, aside, and not-prototypical, other. Taking into account the
characteristics of the subjects through the additional variables enables the fol-
lowing interpretation: the most contributive characteristic at the same time
as the most typical is the membership of the northern hemisphere. In fact
because of the composition of the sample, this one merges with the group of
the trainee teachers of the IUFM of Lyon. These subjects, the more well-read
men of the total sample, by their answers to the Q2 question, clarify their
representations, which are organized around the idea that the “other” located
in the southern hemisphere cannot see the “same moon”. Their training seems
to lead them, through metacognitive distance, to reject the prototypical fig-
ures according to a particular condition which is to be located in the other
hemisphere. This university training specific to teaching professionals could
play a part in this distance as we suggested (2.6.3).
The C3 class incorporated on level 17 present the interesting results: on
the one hand, it is the combined reflection of the rejection of the forms 1B and
1C and the attraction of the forms 1A and 1D; in addition, it arises that its
composition is largely determined by the EchNC_Enf sub-group (Table 6).
 Identifying didactic and sociocultural obstacles to conceptualization 371

This property occurs in the sense that we pointed to leave the analysis of the
similarities. The C4 class incorporated on level 9 is also under strong influ-
ence of the conducts of answers of the subjects CHILDREN (Table 6). This
directed class appears by a source rule (2nao ⇔ 2Anr) which translates the
fact that to answer Q2 negatively a cascade of coherent failures to reply true to
fact, as a logical consequence. This property is almost tautological. However,
what draws our attention is the dominant contributive share of the sub-group
CHILDREN. Indeed, the refusal to believe that a virtual observer can see
these moons in the opposite hemisphere evokes an obstacle of ontogenetic
origin as much as an obstacle of didactic origin.

4 Conclusion
The different symbolic systems which play a role in the process of How has our
methodological reasoning led to the identification of didactic or sociocultural
obstacles? From the point of view of dealing with the statistical data, SIA has
enabled us to establish clear links between binary variables and between class
variables, of which the analysis reveals the importance of the role of didactic
obstacles and of both the school and the extra-school graphic environment.
From this point of view again, the resorting to implicative statistical analysis
with the support of CHIC software enabled us to experience the practical-
ities and to pursue reflection on the issues specifically related to statistical
modelling.
From the point of view of psychology we move away from the central
role of the individual to take into account conceptualization. In this respect,
Bruner [6] wrote: “we were psychologists conditioned by a tradition that put the
individual first. However, the symbolic systems that people use to build mean-
ing are already installed, they are already “there” deeply ingrained in culture
and language”. In this piece of research, we observed that specific symbolic
representations emphasises specific aspects of the concept and gives rise to
potential didactic and/or socio-cultural obstacles. These symbolic representa-
tions are presented in temporal synchrnism with the natural language.
Therefore we spoke here of a whole comprised of linguistic and non lin-
guistic signifiers that are engaged in the teaching-learning situations. These
situations rely on the interaction between the various symbolic systems which
understanding and give rise to specific problems : learning type, conceptual-
ization level and also the nature itself of the concept as described by Vygotski
in his opposition of everyday concepts and scientific concepts. Vergnaud’s the-
ory on conceptual fields taking in consideration the learning context and the
characteristics of the learning method appears to be of greater interest for the
interpretation of our results than the dual opposition proposed by Vygotsky.
Indeed it is clear from our results that school situations and non-school situ-
ations should be differentiated since they bring up distinct focus and level of
awareness.
372 Acioly-Régnier and Régnier

In a school context, the focus is essentially driven toward bipolar rela-

tion [situation↔operational invariant] putting aside the whole referential sit-
uations of life experience. In this case the subject weakness consists in an
inability or difficulty to recognize the situations, not exposed in the school
context or even sometimes developed in this context, where the concept could
be functional. For example: the learners admit that although they know the
definitions they cannot apply them. They give thus the responses concerning
the moon phases that are accepted in the school context, meanwhile their
conceptual level remains quite weak.
On the other hand, in the non school context, the focus is mostly driven
toward the bipolar relation [situation ↔ operational invariant] neglecting thus
the information contained in the signifiers. In this case the subject weakness
consists in the insufficiency of symbolic resources that could allow him to
amplify the knowledge developed locally.
The data which came from the questionnaire corroborates with the obser-
vations made beforehand and described in this article, in the sense that where
the responses seemed overly determined by school learning and by the cultural
environment with the prototypical graphic representations. It must be noted
however that the relationship between images and scientific learning is surely
not established in the school setting. The cultural environment calls for more
and more imagery resources as we have already underlined. Therefore these
data bring to the fore the important role of prototypical figures in the con-
ceptualization of the phases of the moon by well-read subjects. In most cases,
it is not about symbols which represent the sensory experience, but figures
which are taken as a simple recording of images which create an obstacle to
a higher level of conceptualisation.

References
1. N.M. Acioly. A logica matemática no jogo do bicho: compreensão ou utilização
de regras? Master’s thesis, Université Fédérale de Pernambuco, Recife, Brazil,
1985.
2. N.M. Acioly. LA JUSTE MESURE : une étude des compétences mathématiques
des travailleurs de la canne à sucre du Nordeste du Brésil dans le domaine de
la mesure. PhD thesis, Université René Descartes, Paris V, 1994.
3. N.M. Acioly-Régnier. Analyse des compétences mathématiques de publics
adultes peu scolarisés et/ou peu qualifiés. In Illettrismes : quels chemins vers
l’écrit?, Les actes de l’université d’été du 8 au 12 juillet 1996, Lyon, France,
1997. Ed. Magnard.
4. G. Bachelard. La formation de l’esprit scientifique. Paris Lib. Vrin, 1938/1996.
5. G. Brousseau. Les obstacles épistémologiques et les problèmes en mathéma-
tiques. Recherches en Didactique des Mathématiques, 4(2):165–198, 1983.
6. J. Bruner. . . . Car la culture donne forme à l’esprit : de la révolution cognitive
à la psychologie culturell. Paris : Editions Eshel, 1991. Original title: Acts of
Meaning, Harvard University Press.
 Identifying didactic and sociocultural obstacles to conceptualization 373

7. T.N. Carraher, D.W. Carraher, and A.D. Schliemann. Mathematics in the

streets and in schools. British Journal of Developmental Psychology, 3:21–29,
1985.
8. J. Cashford. The Moon Myth and image. London: Cassell Illustrated, 2003.
9. W. Doise and G. Mugny. Le développement social de l’intelligence. Paris :
InterÉditions, 1981.
10. G.G. Granger. Sciences et réalité. Paris : éditions Odile Jacob, 2001.
11. R. Gras. Contribution à l’étude expérimentale et à l’analyse de certaines acqui-
sitions cognitives et de certains objectifs en didactique des mathématiques. PhD
thesis, Université de Rennes 1, 1979.
12. R. Gras, S. Ag Almouloud, M. Bailleul, A. Larher, M. Polo, H. Ratsimba-
Rajohn, and A. Totohasina. L’implication statistique, nouvelle méthode ex-
ploratoire de données. La Pensee Sauvage editions, France, 1996.
13. R. Gras, P. Kuntz, and J.-C. Régnier. Significativité des niveaux d’une hiérarchie
orientée en analyse statistique implicative. Revue des Nouvelles Technologies de
l’Information, RNTI-C-1:39–50, 2004.
14. P. Greenfield and C. Childs. Weaving, color terms and pattern representation
cultural influences and cognitive development among the zinacantecos of south-
ern mexico. International Journal of Psychology, 11:23–48, 1977.
15. E. Harding. Les mystères de la femme : interprétation psychologique de l’âme
féminine d’après les mythes, les légendes et les rêves. Paris : Petite Bibliothèque
Payot, 1953/2001.
16. J. Lave. Cognitive consequences of traditional apprenticeship training. Africa
in Anthropology and Educational Quarterly, 7:177–180, 1977.
17. L. Lebart and A. Salem. Statistique textuelle. Paris Dunod, 1994.
18. G. Lemeignan and A. Weil-Barais. Construire des Concepts en Physique. Paris :
Hachette education, 1993.
19. I.C. Lerman, R. Gras, and H. Rostam. Élaboration d’un indice d’implication
pour les données binaires. Revue Mathématiques et Sciences Humaines, 74 and
75:5–35 and 5–47, 1981.
20. Mc Luhan. Pourquoi des approches interculturelles en sciences de l’éducation.
Bruxelles : De Boeck, 1971/2002.
21. A. Luria. Cognitive Development. Cambridge - MA: Harvard University Press,
1976.
22. G. Mottet. Les situations-images : une approche fonctionnnelle de l’imagerie
dans les apprentissages scientifiques à l’école élémentaire. work document for a
paper in ASTER, n°22, 1996.
23. J. Piaget. Psychologie et pédagogie. Paris : éditions Denoël, 1969.
24. H. Ratsimba-Rajohn. Contribution à l’étude de hiérarchie implicative. Appli-
cation à l’analyse de la gestion didactique des phénomènes d’ostension et de
contradiction. PhD thesis, Université Rennes I, 1992.
25. E.H. Rosch. Cognitive representations of semantic categories. Journal of Ex-
perimental Psychology, 104:192–233, 1975.
26. A.D. Schliemann and N.M. Acioly. Mathematical knowledge developed at work:
the contribution of practice versus the contribution of schooling. Cognition and
Instruction, 6(3):185–221, 1989.
27. S. Scribner. Thinking : reading in cognitive science, chapter Modes of thinking
and ways of speaking : culture and logic reconsidered. London : Cambridge
University Press, 1977.
374 Acioly-Régnier and Régnier

28. D. Toussaint. La lune est-elle menteuse ? bulletin du comité de liaison en-

seignants et astronomes. Les cahiers Clairaut, 87, 1999.
29. G. Vergnaud. Problem solving and concept development in the learning of
mathematics. In E.A.R.L.I. Second Meeting, Tübingen, 1987.
30. G. Vergnaud. Problems of Representation in the teaching and Learning of Math-
ematics, chapter Conclusion chapter. London : Lawrence Erlbaum associates
Publishers, 1987.
31. G. Vergnaud. Psychologie et didactique : quels enseignements théoriques et
méthodologiques pour la recherche en psychologie. In Colloque La Psychologie
Scientifique et ses applications, Clermont-Ferrand, 1987.
32. G. Vergnaud. Questions vives de la psychologie du développement cognitif. In
Colloque d’Aix-en-Provence, 1987.
33. G. Vergnaud. La théorie des champs conceptuels. Recherches en Didactique des
Mathématiques, 10(23):133–170, 1990.
34. G. Vergnaud. Morphismes fondamentaux dans les processus de conceptualisation
— Les Sciences Cognitives en débat. Editions du CNRS, Paris, 1991.
35. L. Vygotski. Pensée et Langage. Paris : Messidor/Editions Sociales, 1985.

Acknowledgments
With thanks to Tim Evans, trainer, for his linguistic skills in his mother-
tongue, and to Pascale Montpied, researcher at the CNRS, whose skills in
English made several useful contributions. Without their help, the English
version of this article would not have been possible.

Appendix 1
NB: The codes of the binary variables are to be found in the questionnaire
[xxx]. The following table is reduced because of space. For Q1, Q2, Q3 and
the first part of Q6, a non-response is coded by the binary variable [xxnr]

Questionnaire

SEX: ( ) M [MALE] ( ) F [FEMALE]

AGE (years ) Profession:
If a teacher, subject taught, and at what level? _______
Previous training: ( ) teacher training college ( ) other. Which? _____
Number of years experience: _______
Place of residence: _______________

Q1.

Have you ever seen the moon like this in reality?

 Identifying didactic and sociocultural obstacles to conceptualization 375

Representations of the moon

Answers
A-Yes A-No B-Yes B-No C-Yes C-No C-Yes C-No
[1Asim] [1Anao] [1Bsim] [1Bnao] [1Csim] [1Cnao] [1Dsim] [1Dnao]

Fig. 13.

Q2.

If someone (from the other hemisphere) tells you that he/she has
never seen any of these moons in reality would you believe that
person? ( ) YES[2sim] ( ) NO [2nao]
If YES: who? ()A[2Asim]/[2Anao] ()B[2Bsim]/[2Bnao] ()C[2Csim]/[2Cnao]
D[2Dsim]/[2Dnao]
Why? If NO, Why not?

Q3.

If someone (from another country but from the same hemisphere)

tells you that he/she has never seen any of these moons in reality,
would you believe that person? ( ) YES[3sim] ( ) NO [3nao]
If YES: who? ( ) A [2Asim]/[2Anao] ( ) B[2Bsim]/[2Bnao] ( ) C[2Csim]/[2Cnao]
( ) D[2Dsim]/[2Dnao]
Why? If NO, Why not?

Q4.

A primary schoolteacher explained his “technique” for teaching the phases of

the moon to his pupils.
He explained that: “the moon tells lies, when it looks like a capital C it is
waning (deCreasing); When it looks like a capital D it is in fact waxing (not
increasing)”.
What do you think of this technique? Is it:( ) adapted [4ADAP] ( ) in-
adapted [4INAD] ( ) efficient [4EFFI] ( ) inefficient [4INEF] Why?

Q5.

A secondary schoolteacher in France explained his “technique” for teaching

the phases of the moon to his pupils. He said: “Its easy! To identify the first
(premier) and last (dernier)quarter of the moon you draw a line Across its
376 Acioly-Régnier and Régnier

diameter. If you get a small letter p (premier) its the first quarter; if you get
the small letter d its the last (dernier) quarter”.
What do you think of this technique? Is it:
( ) adapted [5ADAP] ( ) inadapted [5INAD]
( ) efficient [5EFFI] ( ) inefficient [5INEF] Why?

Fig. 14.

Q6.

A teacher explained that the way the phases of the moon phases are perceived
depends on the hemisphere. For example, it is not seem in the same way from
the southern or from the northern hemisphere. Do you agree with with this
point of view? ( ) YES [6sim] ( ) NO [6nao] Why? If yes: How is the moon seen
from these three places? From the northern hemisphere, the southern hemi-
sphere or the equator. [Use HS (South Hemisphere); HN (North Hemisphere)
or E (Equator) to describe the corresponding representations of the moon.

Representations
of the Moon
[6A_HN] [6B_HN] [6C_HN] [6D_HN]
Hemisphere [6A_E] [6B_E] [6C_E] [6D_E]
[6A_HS] [6B_HS] [6C_HS] [6D_HS]

Fig. 15.
 Identifying didactic and sociocultural obstacles to conceptualization 377

Appendix 2

Sample EchNC_Ad Ech_IUFM Ech_UFRPE EchNC_Enf

198 119 28 29 22
Var. bin. Freq. % Freq. % Freq. % Freq. % Freq. %
[1Asim] 171 86.36 100 84.03 25 89.29 26 89.66 20 90.2
[1Anao] 21 10.61 16 13.45 3 10.71 0 0.00 2 9.09
[1Anr] 6 3.03 3 2.52 0 0.00 3 10.34 0 0.00
[1Bsim] 58 29.29 43 36.13 5 17.86 4 13.79 6 27.3
[1Bnao] 116 58.59 57 47.90 23 82.14 20 68.97 16 72.7
[1Bnr] 24 12.12 19 15.97 0 0.00 5 17.24 0 0.00
[1Csim] 65 32.82 47 36.13 10 35.71 3 10.34 5 22.7
[1Cnao] 110 55.56 54 47.90 18 64.29 21 72.41 17 77.3
[1Cnr] 23 11.62 18 15.97 0 0.00 5 17.24 0 0.00
[1Dsim] 160 80.81 96 80.67 19 67.86 25 86.21 20 90.2
[1Dnao] 25 12.63 17 14.29 6 21.43 0 0.00 2 9.09
[1Dnr] 13 6.57 6 5.04 3 10.71 4 13.79 0 0.00
[2sim] 163 82.32 98 82.35 25 89.29 26 89.66 14 63.6
[2nao] 34 17.17 20 16.81 3 10.71 3 10.34 8 36.4
[2nr] 1 0.51 1 0.84 0 0.00 0 0.00 0 0.00
[2Asim] 40 20.20 25 21.01 10 35.71 3 10.34 2 9.09
[2Anao] 120 60.61 72 60.50 14 50.00 21 72.41 13 59.1
[2Anr] 38 19.19 22 18.49 4 14.29 5 17.24 7 31.8
[2Bsim] 109 55.05 69 57.98 11 39.29 18 62.07 11 50.0
[2Bnao] 48 24.24 27 22.69 11 39.29 6 20.69 4 18.2
[2Bnr] 41 20.71 23 19.33 6 21.43 5 17.24 7 31.8
[2Csim] 112 56.57 72 60.50 13 46.43 15 51.72 12 54.5
[2Cnao] 40 20.20 24 20.17 4 14.29 9 31.03 3 13.6
[2Cnr] 46 23.23 23 19.33 11 39.29 5 17.24 7 31.8
[2Dsim] 54 27.27 33 27.73 11 39.29 6 20.69 4 18.2
[2Dnao] 107 54.04 65 54.62 13 46.43 18 62.07 11 50.0
[2Dnr] 37 18.69 21 17.65 4 14.29 5 17.24 7 31.8
378 Acioly-Régnier and Régnier

Appendix 3

Sample EchNC_Ad Ech_UFRPE EchNC_Enf

170 119 29 22
Var. bin. Freq. % Freq. % Freq. % Freq. %
[3sim] 114 67.06% 74 62.18% 24 82.76% 16 72.73%
[3nao] 51 30.00% 41 34.45% 5 17.24% 5 22.73%
[3nr] 5 2.94% 4 3.36% 0 0.00% 1 4.55%
[3Asim] 31 18.24% 22 18.49% 3 10.34% 6 27.27%
[3Anao] 102 60.00% 76 63.87% 16 55.17% 10 45.45%
[3Anr] 37 21.76% 21 17.65% 10 34.48% 6 27.27%
[3Bsim] 71 41.76% 46 38.66% 16 55.17% 9 40.91%
[3Bnao] 61 35.88% 51 42.86% 3 10.34% 7 31.82%
[3Bnr] 38 22.35% 22 18.49% 10 34.48% 6 27.27%
[3Csim] 65 38.24% 46 38.66% 12 41.38% 7 31.82%
[3Cnao] 67 39.41% 51 42.86% 7 24.14% 9 40.91%
[3Cnr] 38 22.35% 22 18.49% 10 34.48% 6 27.27%
[3Dsim] 30 17.65% 22 18.49% 2 6.90% 6 27.27%
[3Dnao] 102 60.00% 75 63.03% 17 58.62% 10 45.45%
[3Dnr] 38 22.35% 22 18.49% 10 34.48% 6 27.27%
[4ADAP] 42 24.71% 28 23.53% 8 27.59% 6 27.27%
[4INAD] 55 32.35% 49 41.18% 4 13.79% 2 9.09%
[4EFFI] 64 37.65% 34 28.57% 17 58.62% 13 59.09%
[4INEF] 36 21.18% 29 24.37% 2 6.90% 5 22.73%
[5ADAP] 68 40.00% 54 45.38% 7 24.14% 7 31.82%
[5INAD] 40 23.53% 34 28.57% 4 13.79% 2 9.09%
[5EFFI] 80 47.06% 51 42.86% 18 62.07% 11 50.00%
[5INEF] 18 10.59% 11 9.24% 3 10.34% 4 18.18%
[6sim] 111 65.29% 77 64.71% 21 72.41% 13 59.09%
[6nao] 47 27.65% 32 26.89% 8 27.59% 7 31.82%
[6nr] 12 7.06% 10 8.40% 0 0.00% 2 9.09%
[6A_HN] 56 32.94% 50 42.02% 5 17.24% 1 4.55%
[6A_E] 38 22.35% 30 25.21% 6 20.69% 2 9.09%
[6A_HS] 83 48.82% 58 48.74% 17 58.62% 8 36.36%
[6B_HN] 37 21.76% 22 18.49% 10 34.48% 5 22.73%
[6B_E] 35 20.59% 27 22.69% 5 17.24% 3 13.64%
[6B_HS] 25 14.71% 21 17.65% 2 6.90% 2 9.09%
[6C_HN] 27 15.88% 17 14.29% 6 20.69% 4 18.18%
[6C_E] 41 24.12% 30 25.21% 7 24.14% 4 18.18%
[6C_HS] 31 18.24% 23 19.33% 5 17.24% 3 13.64%
[6D_HN] 62 36.47% 52 43.70% 6 20.69% 4 18.18%
[6D_E] 37 21.76% 28 23.53% 5 17.24% 4 18.18%
[6D_HS] 64 37.65% 50 42.02% 12 41.38% 2 9.09%
 Identifying didactic and sociocultural obstacles to conceptualization 379

Appendix 4

Table Appendix 4.1

n = 198 Variables “Questions Q1 & Q2”

GROUPS 1A 1B 1C 1D 2 2A 2B 2C 2D
EchNCAd 10.9 19.9 20.6 11.7 8.15 9.57 6.62 8.64 5.76 χ2
EchIUFM
EchFRPE 0.06 0.10 0.10 0.06 0.04 0.05 0.03 0.04 0.03 ϕ2
EchNCEnf
0.15 0.20 0.21 0.16 0.13 0.14 0.12 0.13 0.11 Tschprow
α = 0.05 I D D I I I I I I
I = INDEPENDENT D = REJECT INDEPENDENCE

Table Appendix 4.2

Question Q1B Question Q1C

Groups 1Bsim 1Bnao 1Bnr 1Csim 1Cnao 1Cnr
EchNC_Ad (A+) (R-) (A+) (A+) (R-) (A+)
Ech_IUFM (R-) (A+) (R-) (A+) (A+) (R-)
Ech_UFRPE (R-) (A+) (A+) (R-) (A+) (A+)
EchNC_Enf (R-) (A+) (R-) (R-) (A+) (R-)
(A+)=Attraction (R-)=Repulsion

Table Appendix 4.3

n = 170 Variables “Questions Q1 & Q2”

GROUPS 1A 1B 1C 1D 2 2A 2B 2C 2D
EchNC_Ad 9.13 10.84 16.00 9.16 6.63 4.91 2.11 3.94 5.76 χ2
Ech_UFRPE 0.05 0.06 0.09 0.05 0.04 0.03 0.01 0.02 0.03 ϕ2
EchNC_Enf
0.16 0.18 0.22 0.16 0.14 0.12 0.08 0.11 0.13 Tschuprow
α = 0.05 I D D I I I I I I

Table Appendix 4.4

n = 170 Variables “Questions Q3 & Q6”

GROUPS 3 3A 3B 3C 3D 6
EchNC_Ad 5.46 6.55 11.5 5.35 7.01 3.00 χ2
EchUFRPE
EchNC_Enf 0.03 0.04 0.07 0.03 0.04 0.02 ϕ2
0.13 0.14 0.18 0.13 0.14 0.09 Tschuprow
α = 0.05 I I D I I I
 Part IV

Extensions to rule interestingness in data

mining
Pitfalls for Categorizations of Objective
Interestingness Measures for Rule Discovery

Einoshin Suzuki

Graduate School of Information Science and Electrical Engineering,

Kyushu University, Japan
[email protected]

Summary. In this paper, we point out four pitfalls for categorizations of objec-
tive interestingness measures for rule discovery. Rule discovery, which is extensively
studied in data mining, suffers from the problem of outputting a huge number of
rules. An objective interestingness measure can be used to estimate the potential
usefulness of a discovered rule based on the given data set thus hopefully serves as a
countermeasure to circumvent this problem. Various measures have been proposed,
resulting systematic attempts for categorizing such measures. We believe that such
attempts are subject to four kinds of pitfalls: data bias, rule bias, expert bias, and
search bias. The main objective of this paper is to issue an alert for the pitfalls
which are harmful to one of the most important research topics in data mining. We
also list desiderata in categorizing objective interestingness measures.

Key words: data bias, rule bias, expert bias, search bias, objective interestingness
measure, rule discovery

1 Introduction

Rule discovery is one of the most extensively studied research topics in data
mining as shown by the proliferation of discovery methods for finding useful
or interesting (i.e. potentially useful) rules [3,5,8,29,34–38]. Such methods can
be classified either objective or subjective [32]. We define that an objective
method evaluates the interestingness of a rule based on the given data set
while a subjective method relies on additional information typically provided
by the user in the form of domain knowledge.
A subjective method often models interestingness more appropriately than
an objective method but cannot be applied to domains where little or no
additional information is available. A subjective method is also prone to
overlooking useful knowledge due to inappropriate use of domain knowledge.
An objective method is free from these problems and poses no cost to the
user for preparing subjective information. In the objective approach, many

E. Suzuki: Pitfalls for Categorizations of Objective Interestingness Measures for Rule

Discovery, Studies in Computational Intelligence (SCI) 127, 383–395 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
384 E. Suzuki

researchers have proposed an objective interestingness measure which re-

turns a real value as an estimated degree of interestingness of a discovered
rule [5, 8, 11, 12, 17, 23, 25, 26, 30, 34].
In general, a categorization of methods brings deep understanding on the
nature of the problem and possibility for further inventions. Because there are
so many objective interestingness measures, a number of papers have tried to
compare, classify, and rank such measures [1, 2, 9, 18, 28, 30, 39]. Each of them
contains interesting proposals and the authors pay attention in interpreting
their experimental results properly. However, many of them seem to give false
conclusions to the readers because all of the categorizations are prone to at
least one of the pitfalls which we warn in this paper.
The quest for interestingness evaluation is undoubtedly important in the
data mining research. For instance, some of the organizers of the ICDM con-
ferences argued that more research for the subject is needed and put interest-
ingness evaluation in one of the ten most important challenges in data mining
(http://www.cs.uvm.edu/~icdm/10Problems/10Problems-05.pdf). In this
paper, we mainly point out data bias, rule bias, expert bias, and search bias as
pitfalls. Our intention is to foster proper research on the interestingness mea-
sures and a sound structuralism in evaluating individual research attempts
for developing a rule discovery method. We also list desiderata in categorizing
objective interestingness measures: respect the true objective, show experi-
mental results in an objective manner, avoid illusion of omnipotent measures,
and be a structuralist in deriving conclusions.
In the rest of this paper, we review objective interestingness measures for
rule discovery and attempts for categorizing the measures in sections 2 and 3,
respectively. We then point out the four pitfalls in section 4. Section 5 gives the
principles to avoid such pitfalls and we try to foster astute readers who would
contribute to a sound development of the data mining community. Section 6
concludes this paper.

2 Objective Interestingness Measures for Rule Discovery

2.1 Rule Discovery

The objective of rule discovery is to obtain a set Π of rules given data D and
additional information α. Here α typically represents domain-specific criteria
such as expected profit or domain knowledge and an element of Π represents
a rule π. As we focus on the objective approach, we assume α = ∅ in this
paper.
The case in which D represents a table, alternatively stated “flat” data, is
most extensively studied in data mining. On the other hand, the case when
data D represent structured data such as time-series data and text data typ-
ically necessitates a procedure for handling such a structure (e.g. [10, 22]).
In order to focus on the interestingness aspect, we limit our attention to the
 Pitfalls for Categorizations of Objective Interestingness Measures 385

former case. In this case, D consists of n examples e1 , e2 , . . . , en . An example

ei is described with m attributes a1 , a2 , . . . , am and an attribute aj takes one
of |aj | values vj,1 , vj,2 , . . . , vj,|aj | . We represent ek = (wk,1 , wk,2 , . . . , wk,m ),
where wk,j ∈ {vj,1 , vj,2 , . . . , vj,|aj | }
A rule represents a local probabilistic tendency in D and can be repre-
sented as A → B. Here A and B are called a premise and a conclusion,
respectively, and each of them specifies a subspace of the example space. For
instance, (a1 = v1,1 ) ∨ (a2 = v2,2 ) → (a3 = v3,1 ) ∧ (a4 = v4,4 ) is a rule. A rule
can be classified into either logical or probabilistic and this paper is concerned
with the latter. A probabilistic rule A → B can have a confidence P (B|A)
smaller than 1 i.e. P (B|A) < 1 while a logical rule necessitates P (B|A) = 1.
Note that here we use each of A and B to represent a set of examples which
reside in the corresponding subspaces. In the rest of this paper, we use this
notation.

2.2 Example of the Objective Interestingness Measures

for Rule Discovery

In computer science, research on objective interestingness measures for rule

discovery goes at least back to 1960’s [6, 17]. Various measures including [8]
have been proposed. Since the objective of this paper is not to conduct a
survey on such measures, we mainly explain the J-measure [34], which exhibits
several desirable properties and is based on [6], as a representative. We leave
explanation on other objective interestingness measures for rule discovery such
as lift, conviction, and intensity of implication to [1, 2, 5, 8, 9, 11, 18, 25, 26, 28,
30, 39]1 .
Consider a rule Y = y → X = x, where the premise Y = y is a single
or a conjunction of “attribute = value”s and the conclusion X = x is an “at-
tribute=value” 2 . Though multiple attributes can be contained in the premise,
we follow the notation of [34] and describe it Y = y, where Y and y corre-
spond to a vector of attributes and a vector of attribute values, respectively.
We define f (X, Y = y) as the instantaneous information that the event Y = y
provides about X, i.e. the information that we receive about X given that
Y = y has occurred. It is shown in [6] that the following j-measure is the only
non-negative function that satisfies Ey [f (X; Y = y)] = I(X; Y ), where Ey (g)
and I(X; Y ) represent the expected value of g in terms of y and the mutual
information between X and Y , respectively.

1
Intensity of implication is an excellent measure which can take the size of the
data set into account. We, however, explain J-measure here because intensity of
implication is explained many times in this book.
2
These assumptions are not necessary but help understanding of a wide range of
readers.
386 E. Suzuki
 
X P (x|y)
j(X; Y = y) = P (x|y) log2 (1)
x
P (x)
   
P (x|y) P (x|y)
= P (x|y) log2 + P (x|y) log2 (2)
P (x) P (x)

(2) is derived due to the nature of rule discovery i.e. it suffices to consider
the events of the conclusion X = x and its negation X = x. In [34], the
interestingness of the rule Y = y → X = x is evaluated with the average
information content of the rule under the name of the J-measure J(X; Y = y).

J(X; Y = y) = P (y)j(X; Y = y) (3)

Note that the J-measure represents the amount of information compressed

by the rule Y = y → X = x: with the rule, the code length for P (x, y) exam-
ples becomes − log2 P (x|y) from − log2 P (x) and the code length for P (x, y)
examples becomes − log2 P (x|y) from − log2 P (x). (3) represents the differ-
ence of the amount of information between the situations with and without
the rule. From the practical viewpoint, a general rule (P (y) high), an accu-
rate rule (P (x|y) high), and a rule predicting a rare event (P (x) low) exhibits
a high J-measure value if other parameters remain unchanged. This nature
makes sense from the viewpoint of discovery of interesting rules with objective
information.
It is worth noting again that a variety of interestingness measures have
been proposed besides the J-measure. They include mutual information
I(X; Y ), support P (x, y), confidence P (x|y), and Jaccard P (x, y)/[P (x) +
P (y) − P (x, y)]. This situation has motivated systematic attempts to catego-
rizing such measures.

3 Categorizations of Objective Interestingness Measures

for Rule Discovery
3.1 Desiderata on Objective Interestingness Measures

In this section, we overview representative categorizations of objective inter-

estingness measures for rule discovery [1, 2, 9, 18, 28, 30, 39]. It should be noted
that Hilderman and Hamilton studied various objective interestingness mea-
sures extensively [13–16]. We omit them in this paper because they mainly
assume that a pattern is represented by a contingency table.
Probably [30] is one of the earliest papers that discuss various objective
interestingness measures for rule discovery. After a brief mention on four mea-
sures, it proposes the following desirable principles for such a measure RI.
1. RI = 0 if P (x, y) = P (x)P (y).
2. RI monotonically increases with P (x, y) when other parameters remain
the same.
 Pitfalls for Categorizations of Objective Interestingness Measures 387

3. RI monotonically decreases with P (x) or P (y) when other parameters

remain the same.
It then proposed nP (x, y) − nP (x)P (y) as the simplest function that satisfies
these principles. It is clear that the author was aware of the limitation of
an objective interestingness measure and just proposed general principles for
such measures.
[39] extends the idea of listing desiderata in [30] and overviews 21 ob-
jective interestingness measures for rule discovery. Based on discussions on
examples of contingency tables of rules, they propose five additional proper-
ties to the above three principles of [30]. Each of the 21 measures is allocated
either “Yes” or “No” for each of the three principles and the five additional
properties. It turns out none of the 21 measures satisfy these 8 criteria and
the authors conclude that there is no measure that is better than others in
all application domains. We believe that the eight criteria are related with
the degree of interestingness of a rule in general but they represent neither
sufficient conditions nor necessary conditions of an accurate interestingness
measure. An objective interestingness measure can serve as a filter to remove
discovered rules which are nearly hopeless from consideration.

3.2 Cluster Analysis of Objective Interestingness Measures

[39] shows an attempt to analyze the 21 objective rule interestingness mea-

sures by ranking 10,000 synthetically generated contingency tables with the
measures and investigating the correlation between pairs of measures. The
results for nine kinds of ranges of support values show that many measures
are highly correlated and can be grouped into a few clusters. We estimate that
the authors are aware of the danger of random experiments and employed a
large number of contingency tables and different ranges for support. It should
be also noted that the authors did not dare to draw drastic conclusions such as
“objective interestingness measures can be classified into five groups” from the
experimental results. We anticipate, however, that a typical reader can mis-
understand their intention and believe the experimental results as evidence of
categorizations of interestingness measures.
[18] goes one step further than [39] and performs clustering instead of cor-
relation analysis. It reports that 34 objective interestingness measures can be
classified into 10 clusters based on their performance on 120,000 association
rules discovered from the Mushroom data set [7]. This attempt is interesting
because it provides empirical results on the categorization of interestingness
measures and some of the results are on their way to be leveraged to empirical
evidence based on persuasive reasons (e.g. a number of interestingness mea-
sures including confidence, Laplace, and causal confidence form a big cluster
because most of them are issued from the confidence measure). Unlike [39],
they seem to use one kind of parameter values thus provide one clustering
result. However, the obvious danger is that even a typical reader might take
388 E. Suzuki

the clustering result as an empirical evidence though it depends on one data

set and one kind of parameter values. We will explain these issues in section 4.

3.3 Incorporating Experts in the Analysis

[28] empirically investigates the performance of 40 objective rule interest-

ingness measures by using a subjective evaluation of a domain expert as
the ground truth. The experiments were performed using a data set on
hepatitis and the domain expert was asked to classify 30 and 21 discov-
ered rules into especially-interesting, interesting, not-understandable, and not-
interesting. The performance of a measure represents how well the measure
estimates the classification of the domain expert. The authors report that re-
call, Jaccard, kappa, CST, χ2 -M, and peculiarity demonstrated the highest
performance. We believe that ranking interestingness measures based on their
performance on sets of discovered rules can give a false conclusion that the
ranking is valid for other rule sets. Subjective evaluation of discovered rules
by domain experts can serve as a ground truth for interestingness measures
but we should be careful not to overgeneralize the conclusions.
[9] investigates effectiveness of objective rule interestingness measures in
estimating the subjective interestingness of a domain expert for each data set
in a more systematic manner than [28]. The paper tries to neutralize the effects
of data sets and rule discovery methods to overcome several shortcomings
of [28]. We will explain these issues in the next section.
The methodologies of [1,2] are more drastic because they depend of classi-
fiers which predict subjective interestingness of a domain expert from results
of objective rule interestingness measures. The training data sets from which
the classifiers are learned contain neither the content of the rules in question
nor domain knowledge. We believe that such attempts are infeasible because
the subjective interestingness of a domain expert depends on both the content
of the rules and domain knowledge.

4 Four Pitfalls to Avoid

4.1 Rule Bias

The categorization depends on the rules to be evaluated. For instance, a mea-

sure called InfoChange-ADT, which requires an exception rule in evaluating
a rule, can not evaluate all rules in [9]. As the authors point out, this failure
was due to the method for discovering the rules: the rules were discovered by
converting decision trees thus had no exception rules. The conclusions derived
in [9] is specific to the method with which the rules are discovered.
As shown in section 3.2, [39] is particularly aware of this pitfall and try
to avoid it by generating 10,000 contingency tables. It is a pity that the
authors give no explanations how they generated these tables as we cannot
 Pitfalls for Categorizations of Objective Interestingness Measures 389

assess the ratio of the generated tables for all possible tables. Anyway, it
is impossible to know the rule set to be evaluated in practice. [9] selects 9
rules out of all rules to be shown to a domain expert for each data set for
investigating the subjective interestingness of the domain expert3 : 3 rules
with the lowest rank, 3 rules in the middle rank, and 3 rules with the highest
rank for each interestingness measure. This method serves as discriminating
3 typical situations but questions remain if it makes sense to choose only 9
rules out of hundreds or thousands of discovered rules. It seems impossible to
avoid rule bias in the categorization but we will show several countermeasures
for this problem in section 5.

4.2 Data Bias

In the categorization research, the rules to be evaluated are typically discov-

ered from data sets. In such a case, the categorization is dependent on the
data sets. For instance, [28] employs a single medical data set. Though the au-
thors emphasize this fact and try to be careful in deriving general conclusions,
any conclusions derived in the paper are specific to the data set. This fact was
pointed out in [9], which clearly shows that the performance of interestingness
measures varies in the eight data sets that they employed.
n
The number of possible table data sets with n binary attributes is Ω(22 )
and the prior probability distribution of the table data sets is unknown. It
seems impossible to avoid data bias in the categorization but we will show
several countermeasures for this problem in section 5.

4.3 Expert Bias

Some of the categorization research ask domain experts to evaluate the degree
of interestingness of discovered rules from their subjective viewpoints and
try to obtain the ground truth of interestingness measures. We agree that
subjective interestingness of domain experts is more important than objective
interestingness in data mining but warn that the results depend on the domain
experts. For instance, [28] asks a single medical expert to evaluate discovered
rules. As the authors point out, any conclusions derived in the paper might
be specific to the domain expert. We have collaborated with medical experts
including the expert in related mining problems and have observed many
discrepancies of opinions among the domain experts [19–21].
It is well-know that domain experts tend to have different opinions for
a non-trivial technical problem, for which rule discovery is expected to be
effective. For instance, [33] explains a situation in which domain experts have
different opinions in classifying volcanoes on Venus and proposes a method for
calculating the minimal probability that domain experts are wrong. It seems

3
We will see this issue in section 4.3.
390 E. Suzuki

that resolving such discrepancies seems impossible: that is why the domain
experts are conducting research.
The authors of [9] point out the weaknesses of [28] and employ eight data
sets in the experiments. They clearly show that the performance of objective
rule interestingness measures varies in eight data sets that they employed.
Moreover, [9] tries to neutralize the effects of rule discovery methods by em-
ploying five classification algorithms. Due to the numerous settings, however,
they could employ one domain expert for each data set and in each case the
domain expert evaluated nine rules (i.e. three rules with the lowest rank, three
rules in the middle rank, and three rules with the highest rank) for each in-
terestingness measure. It is obvious that [9] tolerates some of the problems
of [28] but could not escape from the danger of giving false conclusions to the
reader. We will show several countermeasures for this problem in section 5.

4.4 Search Bias

Some of the categorization research employs a search-based procedure to cate-

gorize objective rule interestingness measures. Typically the results of a search
procedure depends on the values of its parameters and/or its initial conditions.
In such a case, the categorization is dependent on the search procedure.
For instance, as shown in section 3.2, [18] employs a clustering method to
categorize interestingness measures. It is well-known that a typical clustering
method depends on values of its parameters and its initial conditions thus
their results should not be used as a ground truth. In [18], the categoriza-
tion is dependent on the search procedure. As we stated, it seems that they
employ one kind of parameter values and one initial condition. The authors
provide no reasons for seven clusters that they obtained thus these results
can be specific to the problem setting including the parameter values and the
initial condition. The objective of [18] is different and the authors avoid de-
riving general conclusions from the results. We believe, however, some of the
readers might take the results of the clustering as a truth and use it to select
interestingness measures in other cases. We will show several countermeasures
for this problem in section 5.

5 Desiderata for the Categorization Research

5.1 Respect the True Objective

We are aware that most of the papers that we cited as examples of the cat-
egorization research have another objective. For instance, the main objective
of [18] is to develop a tool for selecting the right interestingness measure with
clustering. It is highly recommended to consider the context in which the ex-
periments were performed and not to take a categorization result of objective
rule interestingness measures as a truth. In other words, we should respect the
 Pitfalls for Categorizations of Objective Interestingness Measures 391

context of the paper. However, some of the readers might take a categoriza-
tion result of interestingness measures as a truth. The countermeasure is to
emphasize the true objective the paper and state that a categorization result
represents an example under some conditions.
We have a comment on [3] which is often criticized by researchers in the
“interestingness” community for its two overly simple measures i.e. support
and confidence. The main objective of [3] is to develop fast algorithms for disk-
resident data. As typical researchers in the database community, the authors
assume that the user will select his/her interesting rules with queries thus
support and confidence serve as indexes for a pre-screening. In this sense, [3]
shows another example that we should respect the true objective of a paper.
We are aware of this fact and respect the categorization papers as they have
other objectives.

5.2 Show Experimental Results in an Objective Manner

Any papers in the categorization research should describe how much they
avoided the four biases in showing experimental results in an objective man-
ner. For instance, generating rules to be evaluated randomly is a countermea-
sure for the rule bias. In such a case, the authors should state to what extent
they explored possible kinds of rules under what assumptions (e.g. equal prior
for all rules) they generated the rules. Another typical example is to test
various values for parameters to avoid search bias. For example, a typical
clustering method necessitates a specification of several parameters such as
the number of clusters and a threshold for terminating search. Authors who
employ clustering in a similar manner to [18] should state what kinds of values
they explored under what assumptions.
Data mining as well as related research fields including machine learning
and artificial intelligence cope with ill-structured problems, which have no
clear solution. A typical first step to such kinds of problems is to accumu-
late empirical evidence by analyzing various experimental results then draw
general conclusions. We believe that currently no general empirical evidence
on categorization of objective rule interestingness measures exist despite the
numerous attempts [1, 2, 9, 18, 28, 30, 39].
The results derived in the attempts are special because they heavily de-
pend on at least one of the data sets, discovered rules, domain experts, rule
discovery methods, and analysis methods that they employed. There are at-
tempts to tolerate the effects by employing countermeasures such as random
experiments and multiple methods/experts [9,39] but their effectiveness is lim-
ited due to the huge number of possible choices. We admit that many of the
papers provide reasons for experimental results but we also noticed that such
reasons never consider all the four factors on which the results depend. Sadly
to say, the conclusions derived in the papers such as “34 objective interesting-
ness measures can be classified into 10 clusters” and “recall, Jaccard, kappa,
CST, χ2 -M, and peculiarity demonstrated the highest performance” are not
392 E. Suzuki

guaranteed to be valid in general thus cannot be admitted as an empirical

evidence. We should be particularly cautious against such myths.

5.3 Avoid Illusion of Omnipotent Measures

The objective of machine learning is to realize software which improves its per-
formance by a procedure called learning [27]. We have noticed that a typical
researcher in data mining from the machine learning community is interested
in human factors especially interestingness. At the same time, we are aware
that it is (practically) impossible to realize human intelligence with software
as the proliferation of weak AI shows [31]. Any scientific interest should be
respected but at the same time the reality should be recognized. We should
fight against the illusion that an objective interestingness measure is omnipo-
tent, i.e. it can estimate the subjective interestingness of a discovered rule by
domain experts in general.
Some of the papers [9, 18, 28, 39] try to convey this idea but at the same
time they give the illusion of omnipotence by showing their results on the cate-
gorization. We believe that an objective interestingness measure can just serve
as a naive filter for removing unpromising discovered rules thus should not
be expected to select interesting rules automatically. Classical papers in data
mining [3,30] seem to be aware of the danger and are much more conservative
than recent papers.
As we stated, [3] employs two overly simple measures i.e. support and
confidence, which serve as a measure for a pre-screening. This paper shows
a typical situation that objective rule interestingness measures are used. The
myth of omnipotent measures should be abandoned.

5.4 Be a Structuralist in Deriving Conclusions

A non-structuralist observes behaviors while a structuralist infers their true

cause i.e. the structure of the behaviors. For instance, Claude Levi-Strauss is
said to have discovered that complex taboos on marriage in a tribe can be
clearly explained by considering a woman as a valuable exchange. A sound
criticism on the experimental results is often beneficial. Unexperienced readers
tend to overgeneralize experimental results even to the extent of a truth. In
interpreting experimental results, discussions should include inference on the
true cause.
For instance, it is easy to conclude that objective rule interestingness mea-
sures which put an emphasis on generality are effective from some experimen-
tal results in which such measures often coincide with human interests. The
true reason, however, might be due to the small size of the employed data set,
which resulted in many rules with very low supports. A structuralist approach
is useful in avoiding the pitfalls.
 Pitfalls for Categorizations of Objective Interestingness Measures 393

6 Conclusions

Some of the criticisms might apply to us, as we showed lists of discov-

ered patterns and sometimes asked domain experts to evaluate patterns dis
covered from data sets in our papers such as [36, 38]. However, we carefully
avoided generalizing the results and showed them as experimental results,
and derived conclusions only after we found structural reasons. It should be
also noted that those papers were devoted to proposing discovery methods
for interesting rules and not for categorizing objective rule interestingness
measures.
We believe that in the categorization research the authors should be more
cautious because they evaluate individual efforts for developing discovery
methods for interesting rules. As mentioned in section 1, categorization of
methods is useful and can enhance the quality of the research community.
This paper raises several requirements for a “good” categorization paper based
on four biases which we point out as typical pitfalls. We repeat that each of
the papers of the categorization research has its own objective as mentioned
in section 5.1 and its quality should not be evaluated by the fact that it is
subject to some of the biases.
Some of the researchers might request an operational procedure to cir-
cumvent the pitfalls. We believe that such an operational procedure does not
exist if we employ objective interestingness measures only because interesting-
ness is by its nature subjective. As we have argued, objective interestingness
measures are useful because they serve as filtering out unpromising rules but
are not omnipotent. One of the objectives of this paper is to issue an alert
to beliefs on such an omnipotence. It should be noted that interesting meth-
ods which employ objective interestingness measures as a help in subjective
analysis of experts exist such as [4, 24] and reports on their performance on
real data is demanded.

References
1. H. Abe, S. Tsumoto, M. Ohsaki, and T. Yamaguchi. A Rule Evaluation Support
Method with Learning Models Based on Objective Rule Evaluation Indexes.
In Proc. Fifth IEEE International Conference on Data Mining (ICDM), pages
549–552. 2005.
2. H. Abe, S. Tsumoto, M. Ohsaki, and T. Yamaguchi. Evaluating a Rule Evalua-
tion Support Method Based on Objective Rule Evaluation Indices. In Advances
in Knowledge Discovery and Data Mining (PAKDD), pages 509–519. 2006.
3. R. Agrawal, H. Mannila, R. Srikant, H. Toivonen, and A. I. Verkamo. Fast
Discovery of Association Rules. In Advances in Knowledge Discovery and Data
Mining, pages 307–328. AAAI/MIT Press, Menlo Park, Calif., 1996.
4. J.-P. Barthélemy, A. Legrain, P. Lenca, and B. Vaillant. Aggregation of Valued
Relations Applied to Association Rule Interestingness Measures. In Modeling
Decisions for Artificial Intelligence, LNCS 3885 (MDAI), pages 203–214. 2006.
394 E. Suzuki

5. R. J. Bayardo and R. Agrawal. Mining the Most Interesting Rules. In Proc.

Fifth ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, pages 145–154. 1999.
6. N. M. Blachman. The Amount of Information That y Gives About X. IEEE
Transactions on Information Theory, IT-14(1):27–31, 1968.
7. C. Blake, C. J. Merz, and E. Keogh. UCI Repository of Machine Learning Data-
bases. http://www.ics.uci.edu/~mlearn/ MLRepository.html. Univ. of Calif.
Irvine, Dept. Information and CS (current May 5, 1999).
8. S. Brin, R. Motwani, and C. Silverstein. Beyond Market Baskets: Generalizing
Association Rules to Correlations. In SIGMOD 1997, Proc. ACM SIGMOD
International Conference on Management of Data, pages 265–276. 1997.
9. D. R. Carvalho, A. A. Freitas, and N. F. F. Ebecken. Evaluating the Correlation
Between Objective Rule Interestingness Measures and Real Human Interest. In
Knowledge Discovery in Databases (PKDD), LNCS 3721, pages 453–461. 2005.
10. R. Feldman and I. Dagan. Knowledge Discovery in Textual Databases (KDT). In
Proc. First International Conference on Knowledge Discovery and Data Mining
(KDD), pages 112–117. 1995.
11. R. Gras. Contribution à l’étude expérimentale et à l’analyse de certaines acqui-
sitions cognitives et de certains objectifs didactiques en mathématiques. Thèse
d’Etat, Rennes 1, France, 1979.
12. R. Gras. L’ Implication Statistique. La Pensée Sauvage, 1996. (in French).
13. R. J. Hilderman and H. J. Hamilton. Heuristic for Ranking the Interestingess
of Discovered Knowledge. In Methodologies for Knowledge Discovery and Data
Mining, LNAI 1574 (PAKDD), pages 204–209. 1999.
14. R. J. Hilderman and H. J. Hamilton. Heuristic Measures of Interestingness.
In Principles of Data Mining and Knowledge Discovery, LNAI 1704 (PKDD),
pages 232–241. 1999.
15. R. J. Hilderman and H. J. Hamilton. Applying Objective Interestingness Mea-
sures in Data Mining Systems. In Principles of Data Mining and Knowledge
Discovery, LNAI 1910 (PKDD), pages 432–439. 2000.
16. R. J. Hilderman and H. J. Hamilton. Evaluation of Interestingness Measures
for Ranking Discovered Knowledge. In Knowledge Discovery and Data Mining,
LNCS 2035, pages 247–259. 2001.
17. P. Hájek and C. M. Havel. The GUHA Method of Automatic Hypotheses De-
termination. Computing, 1:293–308, 1966.
18. X.-H. Huynh, F. Guillet, and H. Briand. Clustering Interestingness Measures
with Positive Correlation. In Proc. Seventh International Conference on Enter-
prise Information Systems (ICEIS), pages 248–253. 2005.
19. M. Jumi, M. Ohshima, N. Zhong, H. Yokoi, K. Takabayashi, and E. Suzuki.
Spiral Removal of Exceptional Patients for Mining Chronic Hepatitis Data. New
Generation Computing. (accepted for publication).
20. M. Jumi, E. Suzuki, M. Ohshima, N. Zhong, H. Yokoi, and K. Takabayashi.
Spiral Discovery of a Separate Prediction Model from Chronic Hepatitis Data.
In Proc. Third International Workshop on Active Mining (AM), pages 1–10.
2004.
21. M. Jumi, E. Suzuki, M. Ohshima, N. Zhong, H. Yokoi, and K. Takabayashi.
Multi-strategy Instance Selection in Mining Chronic Hepatitis Data. In Foun-
dations of Intelligent Systems, Lecture Notes in Artificial Intelligence 3488
(ISMIS-2005), pages 475–484. 2005.
 Pitfalls for Categorizations of Objective Interestingness Measures 395

22. E. J. Keogh and M. J. Pazzani. Scaling up Dynamic Time Warping to Massive

Dataset. In Principles of Data Mining and Knowledge Discovery, LNAI 1704
(PKDD), pages 1–11. 1999.
23. Willi Klösgen. Explora: A Multipattern and Multistrategy Discovery Ap-
proach. In Advances in Knowledge Discovery and Data Mining, pages 249–271.
AAAI/MIT Press, Menlo Park, Calif., 1996.
24. P. Lenca, P. Meyer, B. Vaillant, and S. Lallich. On Selecting Interestingness
Measures for Association Rules: User Oriented Description and Multiple Cri-
teria Decision Aid. European Journal of Operational Research. (accepted for
publication).
25. I. C. Lerman. Classification et analyse ordinale des données. Dunod, Paris,
1981.
26. I. C. Lerman, R. Gras, and H. Rostam. Elaboration et évaluation d’un indice
d’implication pour des données binaires. Mathématiques et Sciences Humaines,
74:5–35, 1981.
27. T. M. Mitchell. Machine Learning. McGraw-Hill, Boston, 1997.
28. M. Ohsaki, S. Kitaguchi, K. Okamoto, H. Yokoi, and T. Yamaguchi. Evalua-
tion of Rule Interestingness Measures with a Clinical Dataset on Hepatitis. In
Knowledge Discovery in Databases (PKDD), pages 362–373. 2004.
29. B. Padmanabhan and A. Tuzhilin. A Belief-Driven Method for Discovering
Unexpected Patterns. In Proc. Fourth Int’l Conf. Knowledge Discovery and
Data Mining (KDD), pages 94–100. 1998.
30. G. Piatetsky-Shapiro. Discovery, Analysis, and Presentation of Strong Rules.
In Knowledge Discovery in Databases, pages 229–248. AAAI/MIT Press, Menlo
Park, Calif., 1991.
31. S. Russell and P. Norvig. Artificial Intelligence. Prentice Hall, Upper Saddle
River, New Jersey, 1995.
32. A. Silberschatz and A. Tuzhilin. What Makes Patterns Interesting in Knowledge
Discovery Systems. IEEE Trans. Knowledge and Data Eng., 8(6):970–974, 1996.
33. P. Smyth. Bounds on the Mean Classification Error Rate of Multiple Experts.
Pattern Recognition Letters, 17(12):1253–1257, 1996.
34. P. Smyth and R. M. Goodman. An Information Theoretic Approach to Rule In-
duction from Databases. IEEE Trans. Knowledge and Data Eng., 4(4):301–316,
1992.
35. E. Suzuki. Autonomous Discovery of Reliable Exception Rules. In Proc. Third
Int’l Conf. on Knowledge Discovery and Data Mining (KDD), pages 259–262.
1997.
36. E. Suzuki. Undirected Discovery of Interesting Exception Rules. International
Journal of Pattern Recognition and Artificial Intelligence, 16(8):1065–1086,
2002.
37. E. Suzuki and M. Shimura. Exceptional Knowledge Discovery in Databases
Based on Information Theory. In Proc. Second Int’l Conf. Knowledge Discovery
and Data Mining (KDD), pages 275–278. 1996.
38. E. Suzuki and S. Tsumoto. Evaluating Hypothesis-Driven Exception-Rule Dis-
covery with Medical Data Sets. In Knowledge Discovery and Data Mining, LNAI
1805 (PAKDD), pages 208–211. 2000.
39. P.-N. Tan, V. Kumar, and J. Srivastava. Selecting the Right Interestingness
Measure for Association Patterns. In Proc. Eight ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, pages 32–41. 2002.
Inducing and Evaluating Classification Trees
with Statistical Implicative Criteria

Gilbert Ritschard1 , Vincent Pisetta2 , and Djamel A. Zighed2

1
Dept of Econometrics, University of Geneva, CH-1211 Geneva 4, Switzerland
[email protected]
2
Laboratoire ERIC, University of Lyon 2, C.P.11 F-69676 Bron Cedex, France
{v-pisett, abdelkader.zighed}@univ-lyon2.fr

Summary. Implicative statistics criteria have proven to be valuable interesting-

ness measures for association rules. Here we highlight their interest for classification
trees. We start by showing how Gras’ implication index may be defined for rules
derived from an induced decision tree. This index is especially helpful when the aim
is not classification itself, but characterizing the most typical conditions of a given
conclusion. We show that the index looks like a standardized residual and propose
as alternatives other forms of residuals borrowed from the modeling of contingency
tables. We then consider two main usages of these indexes. The first is purely de-
scriptive and concerns the a posteriori individual evaluation of the classification
rules. The second usage relies upon the strength of implication for assigning the
most appropriate conclusion to each leaf of the induced tree. We demonstrate the
practical usefulness of this statistical implicative view on decision trees through a
full scale real world application.

Key words: Classification tree, Implication strength, Class assignment, Rule rele-
vance, Profile typicality, Targeting

1 Introduction

Implicative statistics was introduced by the French mathematician Régis

Gras [7–9] as a tool for data analysis and has, since the late 90’s, been ex-
ploited for deriving valuable interestingness measures for association rules of
the form “If A is observed, then we are very likely to observe B too” [3,5,10,16].
The basic idea behind implicative statistics is that a statistically observed re-
lationship is of interest only if the number of counter-examples is less than
expected by chance, and that the larger the difference, the more implicative
it is.
We see two major motivations for this concept of statistical implication.
On the one hand, logic implication, does not admit any counter-example.

G. Ritschard et al.: Inducing and Evaluating Classification Trees with Statistical Implicative
Criteria, Studies in Computational Intelligence (SCI) 127, 397–419 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
398 G. Ritschard et al.

Hence, it is too strong and leaves no place for dealing with the random con-
tent of statistical relationships. On the other hand, the classical confidence,
which measures the chances of matching the conclusion when the condition is
satisfied, is not able to tell us whether or not the conclusion is more probable
than it would in case of independence from the condition. For instance, as-
sume that the conclusion B is true for 95% of all the cases. Then, a rule with
a confidence of 90% would do worse than simple chance, i.e. than deciding
that B is true for all cases without taking care of the condition A. But why
looking at counter-examples and not just at positive examples? Indeed, this
is formally equivalent (see Section 2.2), and hence is just a matter of taste.
Looking for the rarity of counter-examples makes the reasoning closer to what
is done with logic rules, i.e. invalidating the rule when there are (too many)
negative examples.
Though, as we will show, this concept of strength of implication is applica-
ble in a straightforward manner to classification rules, only a little attention
has been paid to this appealing idea in the framework of supervised learning.
The aim of this article is to discuss the scope and limits of implicative sta-
tistics for supervised classification and especially for classification trees. One
difference between classification rules and association rules is that the conse-
quent of the former has to be chosen from an a priori set list of classes (the
possible states of the response variable), while the consequent for the latter
can concern any event not involved in the premise, since there is no a priori
outcome variable. A second difference is that unlike the premises of associa-
tion rules, those of a set of classification rules define a partition of the data
set, meaning that there is one and only one rule applicable to each case. These
aspects, however, do not intervene in anyway in the definition of the implica-
tion index which just requires a premise and a consequent. Hence, implication
indexes are technically applicable without restrictions to classification rules.
There remains, nevertheless, the question of whether they make sense in the
supervised learning setting.
The implication index measures how typical the condition of the rule is for
the conclusion, i.e. how much more characteristic than pure chance it is for the
selected conclusion. Indeed, we are only interested in conditions under which
the probability to match the conclusion is higher than the marginal proportion
corresponding to pure chance. A condition with a probability lower than the
marginal proportion would characterize atypical situations for the conclusion,
i.e. situations in which the proportion of cases matching the conclusion is less
than in the whole data set. It would thus be characteristic of the negation
of the conclusion, not the conclusion itself. Looking at typical conditions for
the negation of the conclusion could be useful too. Nevertheless, it does not
require any special attention since it can simply be handled by looking at
the implication strength of the rule in which we would have replaced the
conclusion by its negation.
The information on the gain of performance over chance provided by the
implication index usefully complements the knowledge provided for instance
 Statistical Implicative Criteria for Classification Trees 399

by the classical raw misclassification rate. However, we may go a step further

and, by considering a so called targeting or condition typicality paradigm in-
stead of the classification paradigm, resort to implication indexes for selecting
the conclusion of a rule. Moreover, we could even imagine methods for grow-
ing trees that would optimize the implication strength of the resulting rules.
Such a targeting paradigm will be adopted, for instance, by a physician who
is more interested in knowing the typical profile of persons who develop a
cancer than in predicting for each patient whether or not he has a cancer.
Likewise, a tax-collector may be more interested in characterizing groups in
which he has increased chances to find fakers than in predicting for each tax-
payer whether or not he commits fraud. The most frequent class, commonly
called the ‘majority class’ in the decision tree literature, is obviously the best
choice for minimizing classification errors. However, we will see that for the
targeting paradigm, the highest quality conclusion, i.e. that for which the rule
has the highest implication strength, is not necessarily this majority class.
The paper is organized as follows. Section 2 shows how Gras’ implication
index can be applied to classification rules derived from an induced decision
tree. It proposes alternatives to Gras’ index inspired from residuals used in
the modeling of multiway contingency tables. Section 3 discusses the use of
implication strength for the individual validation of each classification rule. In
Section 4 we adopt the aforementioned typical profile paradigm and consider
using the implication indexes for selecting the most relevant conclusion in a
leaf of a classification tree. We also briefly describe different approaches for
growing trees from that typical profile standpoint. Section 5 reports experi-
mental results that highlight the behavior of the implication strength indexes
and illustrates their potential on a real world application from social sciences.
Finally, we present concluding remarks in Section 6.
We start our presentation by adopting a classical classification standpoint.

2 Classification Trees and Implication Indexes

For our discussion, we consider a fictional example where we are interested in
predicting the civil status (married, single, divorced/widowed) of individuals
from their sex (male, female) and sector of activity (primary, secondary, ter-
tiary). The civil status is the outcome (or response or decision or dependent)
variable, while sex and activity sector are the predictors (or condition or in-
dependent variables). The data set is composed of the 273 cases described by
Table 1.

2.1 Trees and Rules

Classification rules can be induced from data using classification trees in two
steps. First, the tree is grown by seeking, through recursive splits of the learn-
ing data set, some optimal partition of the predictor space for predicting the
400 G. Ritschard et al.

Civil status Sex Activity sector Number of cases

married male primary 50
married male secondary 40
married male tertiary 6
married female primary 0
married female secondary 14
married female tertiary 10
single male primary 5
single male secondary 5
single male tertiary 12
single female primary 50
single female secondary 30
single female tertiary 18
divorced/widowed male primary 5
divorced/widowed male secondary 8
divorced/widowed male tertiary 10
divorced/widowed female primary 6
divorced/widowed female secondary 2
divorced/widowed female tertiary 2
Table 1. The illustrative data set

outcome class. Each split is done according to the values of one predictor. The
process is greedy. It starts by trying all predictors to find the “best” split of
the whole learning data set. Then, the process is repeated at each new node
until some stopping criterion becomes true. In a second step, once the tree
is grown, classification rules are derived by choosing the most relevant value,
usually the majority class (the most frequent), in each leaf (terminal node) of
the tree.
Figure 1 shows the tree induced with the CHAID method [11], using a
5% significance level and a minimal node size fixed at 20. The same tree is
obtained with CART [4] using a minimal .02 gain value. The three numbers in
each node represent the counts of individuals who are respectively ‘married’,

1 2 0 m a r r ie d
0 1 2 0 s in g le
3 3 d iv o r c e d /w id o w e d
m a le fe m a le
s e x
9 6 2 4
1 2 2 2 9 8
2 3 1 0
n o n te r tia r y te r tia r y p r im a r y n o n p r im a r y
s e c to r s e c to r
9 0 6 0 2 4
3 1 0 4 1 2 5 5 0 6 4 8
1 3 1 0 6 4

Fig. 1. Example: Induced tree for civil status (married, single, divorced/widowed)
 Statistical Implicative Criteria for Classification Trees 401

Man Woman
primary or secondary
Civil Status secondary tertiary primary or tertiary Total

Married 90 6 0 24 120
Single 10 12 50 48 120
Div./Widowed 13 10 6 4 33
Total 113 28 56 76 273
Table 2. Table associated to the induced tree

‘single’, and ‘divorced or widowed’. The tree partitions the predictor space
into groups such that the distribution of the outcome variable, the civil status,
differs as much as possible from one group to the other. For our discussion,
it is convenient to represent the four resulting distributions into a table that
cross classifies the outcome variable with the set of profiles (the premises of
the rules) defined by the branches. Table 2 is thus associated to the tree of
Figure 1.
As mentioned, classification rules are usually derived from the tree by
assigning the majority class of the leaf to the branch that leads to it. For
example, a man working in the secondary sector belongs to leaf 3 and will
be classified as married, while a man of the tertiary sector (leaf 4) will be
classified as single. In Table 2, the column headings define the premises of the
rules, the conclusion being given, for each column, by the row containing the
greatest count. Using this approach, the four following rules are derived from
the tree shown in Figure 1:
R1: Man of primary or secondary sector ⇒ married
R2: Man of tertiary sector ⇒ single
R3: Woman of primary sector ⇒ single
R2: Woman of secondary or tertiary sector ⇒ single
In contrast to association rules, classification rules have the following char-
acteristics: i) The conclusions of the rules can only be values (classes) of the
outcome variable, and ii) the premises of the rules are mutually exclusive and
define a partition of the predictor space. Nonetheless, they are rules and we
can then apply to them concepts such as support, confidence and, which is
here our concern, implication indexes.

2.2 Counter-examples and Implication Index

The index of implication (see for instance [6] p 19) of a rule is defined from the
number of counter-examples, i.e. of cases that match the premise but not the
conclusion. In our case, for each leaf (represented by a column in Table 2),
the count of counter-examples is the number of cases that are not in the
majority class. Letting b denote the conclusion (row of the table) of rule j
and nbj the maximum in the jth column, the number of counter-examples
402 G. Ritschard et al.

is nb̄j = n·j − nbj . The index of implication is a standardized form of the

deviation between this number and the number of counter-examples expected
when assuming that the distribution of the outcome values is independent of
the premise.
Formally, the independence hypothesis H0 states that the number Nb̄j of
counter-examples of rule j results from a random draw of n·j cases. Under
H0 , letting nb· /n be the marginal proportion of cases in the conclusion class
b of rule j and setting nb̄· = n − nb· , Nb̄j follows a binomial distribution
Bin(n.j , nb̄· /n), or, when n.j is not fixed a priori, a Poisson distribution with
parameter neb̄j = nb̄· n·j /n [12]. In the latter case, the parameter neb̄j is both
the mathematical expectation E(Nb̄j | H0 ) and the variance var(Nb̄j | H0 ) of
the number of counter-examples under H0 . It is the number of cases in leaf j
that would be counter-examples if they were distributed among the outcome
classes according to the marginal distribution, i.e. that of the root node (right
margin in Table 2).
Gras’ implication index is the difference nb̄j − neb̄j between the observed
and expected numbers of counter-examples, standardized by the standard
deviation, i.e., if we retain the Poisson model,
nb̄j − neb̄j
Imp(j) = q , (1)
neb̄j

which can also be expressed in p terms of the number of cases matching the
rules as Imp(j) = −(nbj − nebj )/ n·j − nebj .
Let us make the calculation of the index explicit for our example. We define
for that the variable “predicted class”, denoted cpred, which takes value 1 for
each case (example) belonging to the majority class of its leaf and 0 otherwise
(counter-example). By cross-classifying this variable with the premises of the
rules, we get Table 3 where the first row gives the number nb̄j of counter-
examples for each rule and the second row the number nbj of examples.
Likewise, Table 4 gives the expected numbers neb̄j and nebj of negative ex-
amples (counter-examples) and positive examples obtained by distributing
the nj· covered cases according to the marginal distribution. Note that these
counts cannot be computed from the margins of Table 3. They are obtained by
first dispatching the column total using the marginal distribution of Table 2
and then separately aggregating each resulting column according to its corre-
sponding observed majority class (not the expected one!). This explains why
Tables 3 and 4 do not have the same right margin.
From these two tables, we can easily get the implication indexes using
formula (1). They are reported in the first row of Table 5. For the first rule,
the index equals Imp(1) = −5.068. This negative value indicates that the
number of observed counter-examples is less than the number expected under
the independence hypothesis, which stresses the relevance of the rule. For the
second rule, the implication index is positive, which tells us that the rule is
 Statistical Implicative Criteria for Classification Trees 403

Man Woman
Predicted class primary or secondary
cpred secondary tertiary primary or tertiary Total
0 (counter-example) 23 16 6 28 73
1 (example) 90 12 50 48 200
Total 113 28 56 76 273
Table 3. Observed numbers nb̄j and nbj of counter-examples and examples

Man Woman
Predicted class primary or secondary
cpred secondary tertiary primary or tertiary Total
0 (counter-example) 63.33 15.69 31.38 42.59 153
1 (example) 49.67 12.31 24.62 33.41 120
Total 113 28 56 76 273
Table 4. Expected numbers neb̄j and nebj of counter-examples and examples

less powerful than pure chance since it generates more counter-examples than
would classifying without taking account of the condition.

2.3 Implication Index and Residuals

In its formulation (1), the implication index looks like a standardized residual,
namely as the (signed square root of) the contribution to the Pearson Chi-
square (see for example [1] p 224). The implication index is indeed related
to the Chi-square that measures the divergence between Tables 3 and 4. The
contributions of each cell to this Chi-square are depicted in Table 5, those of
the first row being the implication indexes.
This interpretation of Gras’ implication index in terms of residuals (resid-
uals for the fitting of the counts of counter-examples by the independence
model) suggests that other forms of residuals used in the framework of the
modeling of the counts in multiway contingency tables could also prove useful
for measuring the strength of rules. These include:

Man Woman
Predicted class primary or secondary
cpred secondary tertiary primary or tertiary
0 (counter-example) -5.068 0.078 -4.531 -2.236
1 (example) 5.722 -0.088 5.116 2.525
Table 5. Contributions to the Chi-square measuring divergence between Tables 3
and 4
404 G. Ritschard et al.
q
The deviance residual , resd (j) = sign(nb̄j − neb̄j ) |2nb̄j log(nb̄j /neb̄j )|, which
is the square root of the contribution (in absolute value) to the likelihood
ratio Chi-square ( [2] pp 136–137).
√ q
Freeman-Tukey’s residual , resF T (j) = nb̄j + 1 + nb̄j − 4neb̄j + 1, which
p

results from a variance-stabilizing transformationq( [2] p 137).

Haberman’s adjusted residual , resa (j) = (nb̄j −neb̄j )/ neb̄j (nb· /n)(1 − n·j /n),
which is the Pearson standardized residual divided by its standard error
( [1] p 224).
There are thus different ways of measuring the departure from the expected
number of counter-examples. It is always instructive to cross-compare values
produced by such alternatives. When they are concordant, as they should be,
comparison reinforces the reliability of the outcome. Divergences, on the other
hand, flag situations for which we should be more cautious before drawing any
conclusion from the numerical value of a given index. Section 5.1 provides some
highlights on the specific behavior of each of the four alternatives considered
here.

Residual Rule R1 Rule R2 Rule R3 Rule R4

Standardized (Gras’ index) ress -5.068 0.078 -4.531 -2.236
Deviance resd -6.826 0.788 -4.456 -4.847
Freeman-Tukey resF T -6.253 0.138 -6.154 -2.414
Adjusted resa -9.985 0.124 -7.666 -3.970
Table 6. The various residuals as alternative implication indexes

Table 6 exhibits the values of these alternative implication indexes for each
of the four rules derived from the tree in Figure 1. We observe that they are
concordant as expected. The standardized residual is known to have a variance
that may be lower than one. This is because the counts nb· and n·j are sample
dependent and hence themselves random. Thus neb̄j is only an estimation of
the Poisson parameter. Ignoring the randomness of the denominator in for-
mula (1) leads to underestimating the strength. The deviance, adjusted and
Freeman-Tukey’s residuals are better suited for this situation and are known
to have in practice a distribution closer to the standard normal N (0, 1) than
the simple standardized residual. We can see in our example that the stan-
dardized residual, i.e. Gras’ implication index, tends to give lower absolute
values than the three alternatives. The only exception is rule R3, for which
the deviance residual provides a slightly smaller value than Gras’ index. Note
that R3 admits only six counter-examples.
 Statistical Implicative Criteria for Classification Trees 405

2.4 Implication Intensity and p-value

In order to evaluate the statistical significance of the computed implication

strength, it is natural to look at the p-value, i.e. at the probability p(Nb̄j ≤
nb̄j | H0 ). When neb̄j is small, this probability can be obtained, conditionally
on nb· and n·j , with the Poisson distribution P (neb̄j ). For large neb̄j , the normal
distribution gives a good approximation. A correction for the continuity may
be necessary, however, because the difference might be for example as large
as 2.6 percent when neb̄j = 100. Letting φ(.) denote the standard normal
 q 
distribution, we have p(Nb̄j ≤ nb̄j | H0 ) ' φ (nb̄j + 0.5 − neb̄j )/ neb̄j .

Residual Rule R1 Rule R2 Rule R3 Rule R4

Standardized (Gras) ress 1.000 0.419 1.000 0.985
Deviance resd 1.000 0.099 1.000 1.000
Freeman-Tukey resF T 1.000 0.350 1.000 0.988
Adjusted resa 1.000 0.373 1.000 1.000
Table 7. The implication intensity and its variants (with continuity correction)

The implication intensity can be defined as the complement of such a

p-value. Gras (see for instance [6]) defines it in terms of the normal approxi-
mation, but without the correction for continuity. We compute it as
 q 
Intens(j) = 1 − φ (nb̄j + 0.5 − neb̄j )/ neb̄j . (2)

In either case, this intensity can be interpreted as the probability of getting,

under the independence hypothesis H0 , a higher number of counter-examples
than the count observed for rule j. Table 7 gives these intensities for our four
rules. It shows also the complement of the p-values of the deviance, adjusted
and Freeman-Tukey’s residuals computed with the continuity correction, i.e.
by adding 0.5 to the observed counts of counter-examples. Notice that pro-
vided probabilities below 50% correspond to positive values of the indexes, i.e.
bad ones, and those above 50% to negative ones. This is a direct consequence
of taking the probabilities from the normal distribution, which is symmetric.

3 Individual Rule Relevance

The implication intensity and its variants are useful for validating each classi-
fication rule individually. This knowledge enriches the usual global validation
of the classifier. For example, among the four rules issued from our illustrative
tree, rules R1, R3 and R4 are clearly relevant, while R2, with an implication
intensity below 50% should be rejected.
406 G. Ritschard et al.

The question is then what shall we do with the cases covered by the con-
ditions of irrelevant rules. Two solutions can be envisaged: i) Merging cases
covered by an irrelevant rule with another rule, or ii) changing the conclusion.
The possible choice of a more suitable conclusion is discussed in Section 4.1.
We exclude indeed further splitting of the node, since we assume that a stop-
ping criterion has been matched. As for the merging of rules, if we want to
respect the tree structure we have indeed to merge cases of a leaf with those
of a sibling leaf, which is equivalent to pruning the corresponding branch. In
our example, this leads to merging rules R1 and R2 into a new rule “Man ⇒
married”. Residuals for the number of counter-examples of this new rules are
respectively ress = −3.8, resd = −7.1, resF T = −4.3 and resa = −8.3. Ex-
cept for the deviance residual, they exhibit a slight deterioration as compared
to the implicative strength of rule R1.
It is interesting here to compare the implicative quality with the error rate
used for validating classification rules. The number of counter-examples con-
sidered is precisely the number of errors produced by the rule on the learning
set. The error rate is thus the percentage of counter-examples among the cases
covered by the rule, i.e. err(j) = nb̄j /n·j , which is also equal to 1 − nbj /n·j ,
the complement to one of the confidence. The error rate suffers that from
the same drawbacks as the confidence. For instance, it does not tell us how
better the rule does than a classification done independently of any condition.
Furthermore, the error rate is linked with the choice of the majority class as
conclusion. For our example, the error rate is respectively for our four rules
0.2, 0.57, 0.11 and 0.36. The second rule is thus also the worst from this point
of view. Comparing with the error rate at the root node, which is 0.56, shows
that this rate of 0.57 is very bad. Thus, for being really informative about the
relevance of the rule, the error rate should be compared with the error rate of
some naive baseline rule. This is exactly what the implication index does. Re-
sorting to implication indexes, we get in addition probabilities which permits
to distinguish between statistically significant and non significant relevance.
Practically, in order to detect over-fitting, error rates are computed on
validation data sets or through cross validation. Indeed, the same can be
done for the implication quality by computing the implication indexes and
intensities in generalization.
Alternatively, we could consider, in the spirit of the BIC (Bayesian in-
formation criteria) or MDL (Minimum message length) principle, to penalize
the implication index by the complexity of the condition. Since the lower the
implication index of a rule j, the better it is, the index should be penalized by
the length kj of the branch that defines the condition of rule j. The general
idea behind such penalization is that the simpler the condition, the lower the
risk to assign a bad distribution to a case. As a first proposal we suggest the
following penalized form inspired from the BIC [14] and based on the deviance
residual
q
Imppen (j) = resd (j) + kj ln(nj ) .
 Statistical Implicative Criteria for Classification Trees 407

For our example, the values of the penalized index are given in Table 3.
These penalized values confirm the ranking of the initial rules, which here
all have the same length kj = 2. In addition, the penalized index is useful for
validating results of merging the two rules R1 and R2. Table 3 highlights the
superiority of the merged rule “Man ⇒ married” over both rules R1 and R2.
It gives a clear signal in favor of merging.
At the root node, both the residual and the number of conditions are zero.
Hence, the penalized implication index is zero too. Thus, a positive penalized
implication index suggests that we can hardly expect that the rule would do
better in generalization than assigning randomly the cases according to the
root node distribution, i.e. independently of any condition. For our example,
this confirms once again the badness of rule R2.

Rule resd k Imppen

R1 -6.826 2 -3.75
R2 0.788 2 3.37
R3 -4.456 2 -1.62
R4 -4.847 2 -1.90
Man ⇒ married -7.119 1 -4.89
Woman ⇒ single -7.271 1 -5.06

Table 8. Implication index penalized for the rule complexity

4 Adopting a Typical Profile Paradigm

To this point, we have assumed that the conclusion of the rule was simply the
majority class. This is justified when the pursued aim is classification. How-
ever, as already mentioned in the introduction, there are situations where the
typical profile paradigm is better suited. Remember the example of the physi-
cian primarily interested in the characteristics of those patients who develop
a cancer, and that of the tax-collector who wants to know the groups of tax
payers who are at most risk of committing fraud. Social sciences, where the
concern is most often to understand phenomena rather than to predict values
or classes, is also a distinctive domain to which the typical profile paradigm
suits well. For example, sociologists of the family may be interested in deter-
mining the profiles in terms of education, professional career, parenthood, etc.
that increase chance of divorce, and in Section 5.2, we present an application
where the goal is to characterize the profiles of those students who are at most
risk of repeating their first year. In such situations, the majority class rule is
no longer the best choice. Indeed, from this typical profile standpoint, it is
more natural to search for rules with the highest possible implication strength
than to minimize the misclassification rate.
408 G. Ritschard et al.

Having this optimal implication strength goal in mind, we successively

discuss the assignment of the most relevant conclusion to the premises defined
by a given grown tree, and the use of implication strength criteria in the tree
growing process.

4.1 Maximal Implication Strength versus Majority Rule

For a given grown tree, maximizing the implication strength is simply achieved
by assigning to each leaf the conclusion for which the rule gets its highest im-
plication intensity. Though ([17] pp 282–287) have already considered this way
of proceeding, they do not provide a sound justification for the approach. Note
also that the method has not, to the best of our knowledge, been implemented
so far in any tree growing software.

Indexes Intensity
Residual married single div./wid. married single div./wid.
Standardized ress 1.6 0.1 -1.3 0.043 0.419 0.891
Deviance resd 3.9 0.8 -3.4 0.000 0.099 0.999
Freeman-Tukey resF T 1.5 0.1 -1.4 0.054 0.398 0.895
Adjusted resa 2.4 0.1 -2.0 0.005 0.379 0.968
Table 9. Implication indexes and intensities of rule R2 for each possible conclusion

To illustrate the principle, we give in Table 9 the values of the alternative

indexes and intensities of implication for each of the three possible conclu-
sions that may be assigned to rule R2 of our example. The conclusion labeled
“single” corresponds to the majority class. However, considering the strength
of implication, the best conclusion is “divorced or widowed”. All four indexes
designate this conclusion as the best with an implication intensity that goes
from 89.1% for Gras’ index to 99.9% for the deviance residual. Indeed, to be a
man working in the tertiary sector is not typical of single people since the rule
would in that case generate more counter-examples than expected by chance.
Concluding to “divorced or widowed” is better in that respect since the number
of positive examples is in that case larger than expected by chance. Again we
can notice that Gras’ index seems to slightly under-estimate the implication
intensity.
An important point is that unlike the majority rule, seeking the maximal
implication strength favors the variability of conclusions among rules, meaning
that we have more chances to create at least one rule for each value of the
outcome variable. In our example, using the majority class we do not create
any rule that concludes with divorced/widowed, while with the implication
strength at least one rule concludes with each of the three outcome states.
Indeed, we need at least as many different profiles as outcome classes if we
want at least one rule concluding with each outcome state, i.e. we should have
r ≤ q with r the number of outcome classes and q the number of rules.
 Statistical Implicative Criteria for Classification Trees 409

By definition, if we assign the same conclusion to all rules, any negative

departure from the expected number of counter-examples of a rule should be
compensated for a positive departure for an other rule. Likewise, for a given
rule, any negative departure from the expected number of counter examples
for one of the possible conclusions should be compensated for a positive one
for an other conclusion. Formally we have
(
e there exists k 6= i such that nk̄j > nek̄j and
nı̄j < nı̄j ⇒
there exists h 6= j such that nı̄h > neı̄h
(
there exists k 6= i such that nk̄j < nek̄j and
nı̄j > neı̄j ⇒
there exists h 6= j such that nı̄h < neı̄h

As a consequence, all the rules cannot attain their maximal implication

strength for the same conclusion, which favors indeed the diversity of the
conclusions among rules. A second consequence is that at each leaf we may
assign a conclusion such that the rule gets a non positive implicative index
or, equivalently, an implication intensity greater or equal to 50%.

4.2 Growing Trees with Implication Strength Criteria

Let us now look at the tree growing procedure and assume that the rule
conclusions are selected so as to maximize the implication strength of the
rules. The question is whether there is a way to split a node so as to maximize
the strength of the resulting rules. The difficulty here is that a split results
indeed in more than one rule. Hence, we face a multicriteria problem, namely
the maximization over sets of implication strengths.
To get simple solutions, one can think to transform the multidimensional
optimization problem into a one dimensional one by focusing on some aggre-
gated criterion. The following are three possibilities:
• A weighted average of the concerned optimal implication indexes, taking
weights proportional to the number of concerned cases.
• The maximum over the strengths of the rules belonging to the set.
• The minimum over the strengths of the rules belonging to the set.
The first criterion is of interest when the goal is to achieve good strengths on
average. The second one should be adopted when we look for a few rules with
high implication strengths without bothering too much for the other ones,
and the latter is of interest when we want the highest possible implication
strength for the poorest rule.
We have not yet experimented tree growing with these criteria. It is worth-
while however to say that, from the typical profile paradigm standpoint meth-
ods such as CHAID that attempt to maximize association seem preferable to
those based on entropies. Indeed, maximizing the strength of association be-
tween the resulting nodes and the outcome variable leads to distributions
410 G. Ritschard et al.

that departure as much as possible from that in the parent node, and hence
from that of the root node corresponding to independence. We may thus ex-
pect the most significant departures from independence and hence rules with
strong implication strength. Methods based on entropy measures, on the other
hand, favor departures from the uniform, or equiprobable, distribution and are
therefore more in line with the classification standpoint.

5 Experimental Results
We present here a series of experimental results that provide additional in-
sights into the behavior and scope of the original implication index and the
three variants we introduced. First, we study the behavior of the indexes.
We then present an application, which also serves as a basis for experimental
investigations regarding the effect of the continuity correction and the conse-
quences of using maximal implication strength rules instead of the majority
rule on classification accuracy, recall and precision.

5.1 Compared Behavior of the 4 Indexes

In order to gain better understanding on how the different implication indexes

behave, we ran a simulation to see how they evolve when the number of
counter-examples is progressively decreased from the expected number under
independence to 0. At independence we expect a null implication strength,
while when no counter-examples are observed we should have high implication
strength.
The simulation design is as follows. We consider a dataset of size 1000 and
a rule defined from a leaf containing 200 cases (20%). We suppose that a pro-
portion p of the 1000 cases belongs to the outcome class selected as conclusion
for the rule. Starting with a proportion f = f0 of cases of the leaf that fall in
the conclusion class, we progressively increase f in 100 constant steps until the
maximum f = 100% is reached. The initial starting point corresponds to inde-
pendence and the final point to a pure distribution with no counter-examples.
At each step we compute, applying the continuity correction, the value of
each of the 4 indexes, namely the standardized, Freeman-Tukey, adjusted and
deviance residuals.
Figure 2 shows the results for p = 10%, 50% and 90%. Notice the difference
of scale between the three plots: The implication strengths are higher when
the class of interest is infrequent in the population, i.e. when p is small. We
observe that the standardized and adjusted residuals evolve linearly between
independence and purity, while the increase in Freeman-Tukey’s residual tends
to accelerate when we approach purity. The deviance residual evolves curiously
in a parabolic way. It dominates the other indexes in the neighborhood of in-
dependence, it reaches a maximum (in absolute terms) and diminishes (in
absolute terms) when we approach purity. This decreasing behavior when the
 Statistical Implicative Criteria for Classification Trees 411

implication index value

-10

standard
-20
F-T
adjusted
-30
deviance

-40

-50
0 20 40 60 80 100
step

(a) 10% of cases in selected outcome class

0
implication index value

standard
F-T
-10
adjusted
deviance

-20
0 20 40 60 80 100
step

(b) 50% of cases in selected outcome class

0
implication index value

standard
F-T
adjusted
deviance

-10
0 20 40 60 80 100
step

(c) 90% of cases in selected outcome class

Fig. 2. Behavior of the 4 indexes between independence (Step 0) and purity (Step
100). Values reported include the continuity correction.
412 G. Ritschard et al.

number of counter-examples tends to 0 disqualifies the deviance residual as

a good measure of the rule implication strength. The linear evolution of the
standadized and adjusted residuals makes them our prefered measures, the lat-
ter having in addition the advantage of being the most reliably comparable
with standard normal thresholds.

5.2 Application on a Student Administrative Dataset

We consider administrative data about the 762 first year students who were
enrolled in fall 1998 at the Faculty of Economic and Social Sciences (ESS)
of the University of Geneva [13]. The goal is to learn rules for predicting the
situation (1. eliminated, 2. repeating first year, 3. passed) of each student
after the first year, or more precisely to discover the typical profile of those
students who are either eliminated or have to repeat their first year. For the
learning data, the response variable is thus the student situation in October
1999. The predictors retained are age, first time registered at University of
Geneva, chosen orientation (Social Sciences or Business and Economics), type
of secondary diploma achieved (classic, latin, scientific, economics, modern,
other), place where secondary diploma was obtained (Geneva, Switzerland
outside Geneva, Abroad), age when secondary diploma was obtained, nation-
ality (Geneva, Swiss except Geneva, Europe, Non Europe) and mother’s living
place (Geneva, Switzerland outside Geneva, Abroad).
Figure 3 shows the tree induced using CHAID with minimal node size set
to 30, minimal parent node size to 50 and a maximal 5% significance for the
Chi-square. Table 10 provides the details regarding the counts in the leafs.
Here, our interest is not in the growing procedure, but rather in the state
assigned to each leaf.

Leaf 6 7 8 9 10 11 12 13 14 Total
1 eliminated 2 17 22 56 31 16 20 18 27 209
2 repeating 1 13 15 48 10 8 16 14 5 130
3 passed 35 87 55 143 28 9 48 12 6 423
Total 38 117 92 247 69 33 84 44 38 762
Table 10. Details about the content of the leafs in Figure 3

Leaf 6 7 8 9 10 11 12 13 14
Majority class 3 3 3 3 1 1 3 1 1
Standardized residual 3 3 3 3 1 1 3 2 1
Freeman-Tukey residual 3 3 3 3 1 1 2 2 1
Deviance residual 3 3 3 2 1 1 2 2 1
Adjusted residual 3 3 3 2 1 1 2 2 1
Table 11. State assigned by the various criteria
percentages.
R o o t
2 7 .4 e lim in a te d
1 7 .1 r e p e a tin g 1 s t y e a r
5 5 .5 p a s s e d
n = 6 7 2

T y p e o f s e c o n d a r y d ip lo m a

E n g in e e r ,A b r o a d ,O th e r C la s s ic ,L a tin ,S c ie n tific E c o n o m ic ,M o d e r n ,M is s in g
1 2 3
4 0 .7 1 6 .6 2 7 .5
2 1 .6 1 1 .7 1 8 .4
3 7 .7 7 1 .7 5 4 .1
n = 1 9 9 n = 2 4 7 n = 3 1 6

N a tio n a lity A g e a t s e c o n d a r y d ip lo m a A g e a t s e c o n d a r y d ip lo m a

S w itz e r la n d , G e n e v a ,
E u ro p e N o n E u ro p e 1 8 o r y o u n g e r 1 9 2 0 o r o ld e r 2 0 o r y o u n g e r 2 1 o r o ld e r
4 5 6 7 8 9 1 0
3 0 .8 5 4 .9 5 .3 1 4 .5 2 3 .9 2 2 .7 4 4 .9
2 0 .5 2 3 .2 2 .6 1 1 .1 1 6 .3 1 9 .4 1 4 .5
4 8 .7 2 1 .9 9 2 .1 7 4 .4 5 9 .8 5 7 .9 4 0 .6
n = 1 1 7 n = 8 2 n = 3 8 n = 1 1 7 n = 9 2 n = 2 4 7 n = 6 9

F ir s t tim e r e g is te r e d O r ie n ta tio n

B u s in e s s a n d S o c ia l
N o Y e s e c o n o m ic s s c ie n c e s
1 1 1 2 1 3 1 4
4 8 .5 2 3 .8 4 0 .9 7 1 .0
2 4 .2 1 9 .1 3 1 .8 1 3 .2
2 7 .3 5 7 .1 2 7 .3 1 5 .8
n = 3 3 n = 8 4 n = 4 4 n = 3 8
Statistical Implicative Criteria for Classification Trees

top to down: eliminated, repeating 1st year, passed. Figures next to the bars are
Fig. 3. CHAID induced tree for the ESS Student data. Outcome states are from
413
414 G. Ritschard et al.

We used successively the majority class rule and each of the four variants
of implication indexes for that. Table 11 reports the results. We can see that
the 5 methods agree for 6 out of the 9 leaves. The conclusion assigned to
leaves number 9, 12 and 13 vary, however, among the 5 methods. All four
implication indexes assign state 2, “repeating the first year”, to leaf 13 where
the majority class is 1, “eliminated”. This tells us that belonging to this leaf,
i.e having a not typical Swiss college secondary diploma obtained either in
Geneva or abroad and having chosen a business and economic orientation, is
a typical profile of those who repeat their first year. And this holds, indeed,
despite “repeating the first year” is not the majority class of the leaf.
The deviance and adjusted residuals agree about assigning also state 2,
“repeating”, to leaves number 9 and 12, and the Freeman-Tukey residual agrees
also with this conclusion for leaf 12. These leaves also define characteristic
profiles of those who repeat their first year, even though the majority class
for these profiles is “passed”.

5.3 Effect of Continuity Correction

We expect continuity correction, i.e. adding .5 to the observed counts nb̄j of

counter-examples, to have only very marginal effects and to be important only
in conjunction with small minimal node sizes.
For our application on the ESS student data, the continuity correction
changes the conclusion only when we use the Freeman-Tukey residual for leaf
12 (with 84 cases). The conclusions remain the same for all other leaves and
for all leaves when we use any of the three other residuals. Furthermore, the
effect of the continuity correction vanishes when we multiply all the counts
by a factor greater or equal to 1.4, which confirms our expectation.
Nevertheless, we suggest systematically introducing the error correction
when computing the indexes. There are two reasons for this: First, it does
not change much the index values in case of large counts and produces values
best suited for comparison with standard normal thresholds in case of small
counts. Secondly, it avoids possible troubles (division by zero for instance)
that may occur when some observed counts are zero.

5.4 Recall and Precision

In terms of the overall error rate, selecting the majority class is no doubt
the better choice. However, if we are interested in the recall rate, i.e. in
the proportion of cases with a given output value ck that are detected as
having this value, we may expect the implication indexes to outperform the
majority rule for infrequent classes. Indeed, highly infrequent outcome states
have high chances to never be selected as conclusion by the majority rule. We
may therefore expect low recall for them when we select the most frequent
class as conclusion. Regarding precision, i.e. the proportion of cases classified
as having a value ck that effectively have this value, expectations are less
 Statistical Implicative Criteria for Classification Trees 415

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
majority standard adjusted deviance FT

Fig. 4. Correct classification rate, 10-fold CV

clear since the relationship between the numerator and denominator seems
not linked to the way of choosing the conclusion.
In order to verify these expectations on our ESS student data, we computed
for the majority rule and each of the four variants of implication indexes, the
10-fold cross-validation (CV) values of the overall good classification rate, as
well as of the recall and precision for each of the three outcome states. As can
be shown on Figure 4 the loss in accuracy that results from using maximal
implication rules lies between 12% for the adjusted residual and 10% for the
standard residual.
Figure 5 exhibits the CV recall rates obtained for each of the three states.
They confirm our expectations: selecting the conclusion according to implica-
tion indexes deteriorates the recall for the majority class “passed”, but results

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0

majority standard adjusted deviance FT majority standard adjusted deviance FT

(a) passed (b) repeating

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
majority standard adjusted deviance FT

(c) eliminated

Fig. 5. Recall, 10-fold CV

416 G. Ritschard et al.

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
majority standard adjusted deviance FT majority standard adjusted deviance FT

(a) passed (b) repeating

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
majority standard adjusted deviance FT

(c) eliminated

Fig. 6. Precision, 10-fold CV

in an improvement in recall for the two other classes. The improvement is

especially important for the last frequent state, i.e. “repeating”, for which we
get recall rates ranging between 30% and 40% instead of almost 0% with the
majority rule.
In Figure 6, we observe an improvement in precision for “passed” (the ma-
jority class) and “repeating” (the last frequent class) and a slight deterioration
for “eliminated”. This illustrates that the choice of the conclusion has appar-
ently no predictable effect on precision. Indeed, the only thing we may notice
here is that improvement concerns the two classes with a proportion of cases
that is further (on either side) from the equiprobable probability 1/c, where
c is the number of outcome classes.

6 Conclusion

The aim of this article was to demonstrate the usefulness of the concept of im-
plication strength for rules derived from induced decision trees. We have shown
that Gras’ implication index can be applied in a straightforward manner to
classification rules and have proposed three alternatives inspired from resid-
uals used in the statistical modeling of multiway contingency tables, namely
the deviance, adjusted and Freeman-Tukey residuals. As for the scope of the
implication indexes we have successively discussed their use for evaluating
individual rules, for selecting the conclusion of the rule and as criteria for
 Statistical Implicative Criteria for Classification Trees 417

growing trees. We have stressed that implication indexes are a valuable com-
plement to classical error rates as validation tools. They are especially inter-
esting in a targeting framework where the aim is to determine the typical
profile that leads to a conclusion rather than classifying individual cases. As
criteria for selecting the conclusion, they may be a useful alternative to the
majority rule in the case of imbalanced data. Their advantage is that in such
imbalanced situation and unlike decisions based on the majority class, they
favor conclusion diversity among rules as well as recall for poorly represented
classes.
Four variants of implication indexes have been discussed. Which one should
we use? The simulation study of their behavior has shown that the deviance
residual curiously diminishes when the number of counter-examples tends to
zero and should therefore be disregarded. The standard residual (Gras’ in-
dex) and Haberman’s adjusted residual both evolve linearly between indepen-
dence and purity and thus seem to be the better choices. From the theoretical
standpoint, if we want to compare the values with thresholds of the standard
normal, Haberman’s adjusted residual is preferable.
We have also introduced the implication intensity as the probability to
get by chance more counter-examples than observed. This is indeed just a
monotonic transformation of the corresponding implication index. Hence rank-
ings based on the indexes or on the intensities will necessarily agree. Indexes
seem better suited, however, to distinguishing between situations with high
implication strengths. The intensities on the other hand, provide additional
information about the statistical significance of the implication strength.
It is worth mentioning that, to our knowledge, implication indexes have not
so far been implemented in tree growing software. Making them available is
essential for popularizing them. We have begun working on implementing the
maximal implication selection process and tree growing algorithms based on
implication criteria into Tanagra [15] a free open source data mining software,
and plan also to make these tools available in Weka.
Beside this implementation task, there are some other issues that would
merit further investigation. For instance, the penalized implication index we
proposed in Section 3 is not completely satisfactory. In a n-arry tree the paths
to the leaves are usually shorter than in a binary tree, even if they define the
same leaves. Penalization based on the length of the path as we proposed,
would therefore be different for a rule derived from a binary tree than for the
same rule derived from a n-arry tree. The use of implication criteria in the
tree growing process needs also a deeper reflection.
Despite all which remains to be done, our hope is that this article will
contribute to enlarge both the scope of induced decision trees and that of
implication statistics.
418 G. Ritschard et al.

References
1. Alan Agresti. Categorical Data Analysis. Wiley, New York, 1990.
2. Yvonne M. M. Bishop, Stephen E. Fienberg, and Paul W. Holland. Discrete
Multivariate Analysis. MIT Press, Cambridge MA, 1975.
3. Julien Blanchard, Fabrice Guillet, Régis Gras, and Henri Briand. Using
information-theoretic measures to assess association rule interestingness. In
Proceedings of the 5th IEEE International Conference on Data Mining (ICDM
2005), pages 66–73. IEEE Computer Society, 2005.
4. L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone. Classification And
Regression Trees. Chapman and Hall, New York, 1984.
5. Henri Briand, Laurent Fleury, Régis Gras, Yann Masson, and Jacques Philippe.
A statistical measure of rules strength for machine learning. In Proceedings
of the Second World Conference on the Fundamentals of Artificial Intelligence
(WOCFAI 1995), pages 51–62, Paris, 1995. Angkor.
6. R. Gras, R. Couturier, J. Blanchard, H. Briand, P. Kuntz, and P. Peter.
Quelques critères pour une mesure de qualité de règles d’association. Revue
des nouvelles technologies de l’information RNTI, E-1:3–30, 2004.
7. R. Gras and A. Larher. L’implication statistique, une nouvelle méthode
d’analyse de données. Mathématique, Informatique et Sciences Humaines,
(120):5–31, 1992.
8. R. Gras and H. Ratsima-Rajohn. L’implication statistique, une nouvelle méth-
ode d’analyse de données. RAIRO Recherche Opérationnelle, 30(3):217–232,
1996.
9. Régis Gras. Contribution à l’étude expérimentale et à l’analyse de certaines ac-
quisitions cognitives et de certains objectifs didactiques. Thèse d’état, Université
de Rennes 1, France, 1979.
10. Sylvie Guillaume, Fabrice Guillet, and Jacques Philippe. Improving the discov-
ery of association rules with intensity of implication. In Jan M. Zytkow and
Mohamed Quafafou, editors, Proceedings of the Eurpoean Conference on Prin-
ciples of Data Mining and Knowledge Discovery (PKDD 1998), volume 1510 of
Lecture Notes in Computer Science, pages 318–327. Springer, 1998.
11. G. V. Kass. An exploratory technique for investigating large quantities of cate-
gorical data. Applied Statistics, 29(2):119–127, 1980.
12. I. C. Lerman, R. Gras, and H. Rostam. Elaboration d’un indice d’implication
pour données binaires I. Mathématiques et sciences humaines, (74):5–35, 1981.
13. Claire Petroff, Anne-Marie Bettex, and Andràs Korffy. Itinéraires d’étudiants
à la Faculté des sciences économiques et sociales: le premier cycle. Technical
report, Université de Genève, Faculté SES, Juin 2001.
14. Adrian E. Raftery. Bayesian model selection in social research. In P. Marsden,
editor, Sociological Methodology, pages 111–163. The American Sociological As-
sociation, Washington, DC, 1995.
15. Ricco Rakotomalala. Tanagra : un logiciel gratuit pour l’enseignement et la
recherche. In Suzanne Pinson and Nicole Vincent, editors, Extraction et Gestion
des Connaissances (EGC 2005), volume E-3 of Revue des nouvelles technologies
de l’information RNTI, pages 697–702. Cépaduès, 2005.
16. Einoshin Suzuki and Yves Kodratoff. Discovery of surprising exception rules
based on intensity of implication. In Jan M. Zytkow and Mohamed Quafafou,
editors, Principles of Data Mining and Knowledge Discovery, Second European
 Statistical Implicative Criteria for Classification Trees 419

Symposium, PKDD ’98, Nantes, France, September 23-26, Proceedings, pages

10–18. Springer, Berlin, 1998.
17. Djamel A. Zighed and Ricco Rakotomalala. Graphes d’induction: apprentissage
et data mining. Hermes Science Publications, Paris, 2000.
On the behavior of the generalizations
of the intensity of implication:
A data-driven comparative study

Benoît Vaillant1 , Stéphane Lallich2 , and Philippe Lenca3

1
IUT de Vannes, Université de Bretagne Sud, VALORIA, France
[email protected]
2
Université Lyon 2, Equipe de Recherche en Ingénierie des Connaissances, France
[email protected]
3
Institut TELECOM, TELECOM Bretagne, Lab-STICC, France
[email protected]

Summary. In this chapter, we present an original and synthetical overview of most

of the commonly used association rule statistical interestingness measures intro-
duced in previous works. These measures usually relate the confidence of a rule to
an independency reference situation. Others relate it to indetermination, or impose
a minimum confidence threshold. We propose a systematic generalization of these
measures, taking into account a reference point, chosen by an expert, in order to
apprehend the confidence of a rule. This generalization introduces new connections
between measures. They lead to the enhancement of some measures. We then pro-
pose new parameterized possibilities. The behavior of the parameterized measures is
illustrated using classical datasets, and these measures are compared to their origi-
nal counter-parts. This study highlights the different properties of each of them and
discusses the advantages of our proposition.

Key words: Statistical interestingness measures, intensity of implication, general-

ized measures.

This generalization introduces new connections between measures. They

lead to the enhancement of some measures. We then propose new parameter-
ized possibilities.

1 Introduction
In this chapter, we focus on the generalization of statistical interestingness
measures. We will consider objective association rule interestingness measures,
which aim at quantifying the quality of rules extracted from binary transac-
tional datasets. Such measures are said to be objective since they only rely

B. Vaillant et al.: On the behavior of the generalizations of the intensity of implication: A data-
driven comparative study, Studies in Computational Intelligence (SCI) 127, 421–447 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
422 B. Vaillant et al.

on frequency counts on the data in order to assess the interest of a rule, as

opposed to subjective ones which are based on expressed prior knowledge. An
association rule is an implication A → B, where A and B (also called itemsets)
are conjunctions of attributes. We denote by n the total number of trans-
actions in the database, na (resp. nb , nab , nab̄ ) the number of transactions
matching A (resp. B, A and B, A but not B), and by pa (resp. pb , pab , pab̄ )
the corresponding relative frequencies. Most measures are expressed as real
valued functions of n, of the marginal frequencies pa , pb , and either pab or pab̄ ,
i.e. as functions of n, and of the confidence pab /pa and marginal frequency
counts of the considered rule since pab̄ =pa -pab . Considering that the more
counter-examples to a rule there are, the worse it is, we restrict our set of
measures to those decreasing with pab̄ , pa and pb being fixed.
Support (Sup) and confidence (Conf) are the most famous of such mea-
sures, being the fundamental principles of Apriori-like algorithms [1]. These
algorithms extract rules such that their Sup and Conf are above given con-
stant thresholds, σs and σc . They are deterministic [9], and produce a large
number of differing rules (see Figure 1, presenting the distribution of Sup and
Conf values of 15771 automatically extracted rules on the Flag dataset),
which, moreover, may not be interesting:
• an expert end-user expects from a rule that its Conf should be above
a reference value, but this reference value seldom if ever equals σc . In
this context, two main lower references are clearly identified as worthy
from a user point of view. The first one is pb , which corresponds to the
independence of the itemsets A and B [33]. In this case the user will focus
on rules such that the prior knowledge of A increases the knowledge of B.
An alternative reference sometimes used is 0.5, as in [4]. In our opinion, the
first reference is to be taken within a targeting strategy, and the second one
when considering a predictive strategy. For example, let us consider an item
B, corresponding to a given kind of cancer whose a priori probability is pb =
0.02, and a rule A → B such that pb|a = 0.20. This rule is very interesting
in a targeting strategy (which aims at identifying risked groups), as the
group of individuals having the characteristics of the itemset A have ten
times more chance of developing the considered cancer than usual. On the
other hand, the rule A → B is not interesting from a predictive point of
view since an individual presenting the characteristics of the itemset A has
a risk far less than 0.5 of developing a cancer. More generally, a user may
be interested in taking into account a rule dependant reference value θ,
0 < θ ≤ 1, and will consider only rules having a Conf greater than θ [25].
• what is more, the data mined is often subject to some sampling scheme. In
order to take that into account, a special kind of measures has been pro-
posed. They are called “statistical” in the sense that, unlike “descriptive”
measures, their value rises with n, the relative frequencies being fixed. Let
us consider the rule A → B, for which pa = 0.30, pb = 0.50, pab = 0.25 and
two situations: n = 20 and n = 400. The value of the correlation coefficient
 Generalized intensity of implication 423

r between A and B being the same (r = 0.22) for each alternative, r is a

descriptive measure. Another possible interestingness measure is the com-
plement to 1 of the p-value of r, denoted by M = 1 − P (r > robs ), where
robs is the observed value of r. In this case, M rises from M = 0.835 for
n = 20 up to M = 0.999 for n = 400. M is thus a statistical measure,
and more precisely a probabilistic one (see section 2.1). This consideration
accounts for developing an inferential approach, and retaining only rules
that are significantly well evaluated by measures, in comparison to the
reference chosen. Amongst the issues that arise from this approach, vali-
dating a large number of rules through the control of false rules discovery
is assessed in [23].

Fig. 1. Sup and Conf values of the rules extracted from the Flag database

Figure 2 illustrates these various conflicting situations. We represent the

Sup value of two families of rules. The first family, denoted by r1 , is defined
such that pa = 0.2 and pb = 0.4. For the second family, r2 , we impose that
pa = 0.45 and pb = 0.8. Given such characteristics, clearly, the equilibrium
and independence situations appear in a different order for r1 and r2 when
pab̄ increases. What is more, if one tries to discard uninteresting rules using a
Sup threshold, either all interesting rules from r1 will be discarded if this σs
is fixed considering r2 , or many uninteresting rules from r2 will be retained if
one considers r1 as reference [39].
As previously mentioned, APRIORI-like algorithms may produce huge
numbers of rules, and thus an essential step in association rule mining is
424 B. Vaillant et al.

Fig. 2. Variations of Sup for two families of rules

the evaluation of their interestingness. The support and confidence frame-

work is not satisfactory, and many new measures have been proposed. Each
new measure is supposed to better highlight a user desired kind of knowledge.
Various properties of interestingness measures -for various data mining tasks-
have been investigated, in particular in [8, 12, 15, 16, 20, 24, 28, 32, 33, 36, 38].
In [29] we propose an extensive study of twenty well-known association rule
interestingness measures based on eight user-oriented points of view.
One of these properties is related to the reference value to which the mea-
sure compares confidence, this reference value being commonly either pb (in-
dependence), or 0.5 (indetermination). We here extend this concept to a user
chosen reference.
This work is concerned with implicative statistical analysis, in which one
tries to measure the strength of a rule A → B. The basic idea behind this
analysis is that the fewer counter-examples observed (i.e. in the data) there
are, the more implicative the rule A → B is.
Since its origins –binary data, mathematical didactics situations [10, 31]–
the implicative statistical analysis has achieved many developments and ap-
plications in various data ming tasks (see for example [7] and [11] for recent
reviews): treatment of modal variables [2], numerical ones [19] and ordinal
ones [14]; user-driven process for mining association rules [18], for classifica-
tion trees [34] and [35], exception rules mining [37], classification association
rules [17].
Objective association rule interestingness measures usually compare the
confidence of a rule to a reference value corresponding to the independence
 Generalized intensity of implication 425

between the antecedent and the consequent of a rule. Some of them com-
pare the confidence of a rule to 0.5, which corresponds to an indetermination
situation.
In a previous work, [25], we first suggested to parameterizing the reference
value of both descriptive and statistical measures in order to compare the
confidence to a reference value θ chosen by the user. The case of statistical
measures, especially the intensity of implication and its generalizations has
been explored in [21]. Theoretical aspects of our works were extended in [26].
In this chapter, we present our results on generalized statistical measures
and we propose an original data driven comparative study of the behavior of
generalized statistical measures.
This chapter is organized as follows. In section 2, we present a general
synthetic overview of statistical measures making reference to independence:
modeling of counter-examples distribution, construction of statistical and
probabilistic measure, enhancement of the discriminating power of the sta-
tistical measures. We introduce in section 3 the statistical measures making
reference to indetermination. Section 4 deals with the generalization of sta-
tistical measures. Discriminant versions of generalized measures are proposed
in section 5. Finally, we conclude in section 6.

2 Statistical measures making reference to independence

2.1 Characteristics of statistical and probabilistic measures

A statistical measure evaluates how far the observed rule is from a null hy-
pothesis H0 corresponding to a lower reference point. From the definition of
a statistical measure, which is a modeling of the kind of rules that one wishes
to discover, it is then possible to define a probabilistic measure as the proba-
bility of obtaining a value of the statistical measure, at most equal to what is
observed, given that the null hypothesis H0 is true.
Classically, this null hypothesis is the hypothesis of independence between
itemsets A and B, and it is tested against a one sided alternative hypothesis
H1 of positive dependence. The corresponding test can be written in terms
of theoretical frequencies referring to A and B (π(·) being the theoretical fre-
quencies):

H0 : πb/a ≤ πb against H1 : πb/a > πb

The modeling of the null hypothesis of independence performed in [31] can
be done in three different ways, with respectively 1, 2 and 3 hazard levels.
0
In [25] we proposed an alternative to the first modeling, denoted 1 . The
four modelings are synthesized in Table 1.
We denote by Nab the random variable generating nab , and H, B and P oi
refer to the hypergeometric, binomial and Poisson distributions, respectively.
426 B. Vaillant et al.

• Modeling 1 (only one hazard level): margins are fixed, only the joint ab-
solute frequencies are random, but with only one degree of freedom.
– The modeling proposed by [31] applies to the distribution of examples,
within the 4 inner possibilities of the contingency table of (A, B), na
and nb being fixed, following a traditional statistical process.
Under H0 , Nab here follows the hypergeometric law H(n, na , pb ). Test-
ing H0 thus means testing the equality of the theoretical confidence of
A → B and A → B, at fixed margins.
• Modeling 10 (only one hazard level): still at fixed margins, an alterna-
tive approach which only takes into account the distribution of examples
between AB and AB is proposed in [25].
– In this case, Nab follows the binomial law B(na , pb ). Testing H0 then
means testing the conformity of the theoretical confidence of A → B, pb
being fixed beforehand.
• Modeling 2 (two hazard levels): modeling 2 of [31] corresponds to modeling
10 , with na also randomized.
– On a first hazard level, it is thus here assumed that Na follows the
binomial law B(n, pa ).
– On a second hazard level, conditionally to Na = na , Nab follows the
binomial law B(na , pb ). Thus Nab follows the binomial law B(na , pa pb ).
• Modeling 3 (three hazard levels): modeling 3 of [31] once again relies on
modeling 10 , where the values of na , and then n are successively random-
ized.
– On the first hazard level, N is assumed to follow the Poisson law
P oi(n).
– On the second hazard level, it is assumed that Na follows the binomial
law B(n, pa ), conditionally to N = n.
– On the third hazard level, and conditionally to N = n and Na = na ,
it is assumed that Nab follows the binomial law B(na , pb ). In this case,
Nab follows the Poisson law P oi(npa pb ).
The statistical and probabilistic measures based on Nab̄ are built as follows:
• by establishing the law of Nab and Nab under the null hypothesis (H0 ) fol-
lowing the chosen modeling, we can express a centered and reduced index4
under H0 (CR notation). In order to have a decreasing quality measure with
respect to nab̄ , the statistical index is defined by SI(i) = −Nab
CR
, where i
refers to the corresponding modeling.
• under standard conditions, the law of this index can be approximated
by the normal distribution, leading to the definition of a probabilistic
measure, defined as the complement to 1 of the surprise of observing
such an exceptional value of the index under H0 . This probabilistic index

4 x−µ
Given a random variable X, its centered and reduced expression is xCR = √ ,
v
where µ is the mean of X and v its variance.
 Generalized intensity of implication 427

is denoted by PI(i) = P (N (0, 1) > nCR

ab
), where i again refers to one of
the four modelings introduced.
The chosen modeling does not affect the expectation, but does modify
the variance. [13] and [6] prefer the third modeling, which dissociates most
the rules A → B and B → A, whereas the first modeling makes no distinction
between these rules. The probabilistic measure hence obtained is the intensity
of implication (IntImp = PI(3) ), which satisfies many properties one expects
a measure should have [12, 28].

Fig. 3. Evolution of IntImp in function of -ImpInd

√
The statistical measure obtained with modeling 1 is SI (1) = r n. The
√
corresponding probabilistic measure PI(1) = P (N (0, 1) < r n) is the com-
plement to 1 of the p-value of r. It is to be noted that in the boolean case,
nr2 = χ2 . Hence PI(1) is the unilateral counterpart of the complement to 1
associated with the χ2 test of independency.
Figure 3 shows that IntImp is an anamorphosis of -ImpInd ([13], see SI (3)
in Table 1) through the normal distribution function.

2.2 Discriminating power of statistical measures

Although it has many good properties [27, 29], one of the major drawbacks of
IntImp (drawback shared by the other statistical and probabilistic measures)
is the loss of discriminating power: by its definition, it will evaluate rules sig-
nificantly different from independence between 0.95 and 1. If n becomes large,
which is particularly true in a data mining context, the slightest divergence
from an independence situation becomes highly significant, thus leading to
high and homogeneous values of the measure, close to 1. It is thus difficult to
428 B. Vaillant et al.

select the best rules. For example, we computed the values taken by IntImp
on rules extracted from three classical datasets [40]. The Breast Cancer, Con-
traceptive Method Choice and Housing datasets are available from the UCI
repository (http://www.ics.uci.edu/~mlearn/databases/).
On the Breast Cancer data, containing n = 683 entries, 3079 rules have
an IntImp value above 0.99, out of the 3095 rules generated by Apriori,
with σs = 0.10 and σc = 0.70. On the Contraceptive Method Choice data,
containing n = 1473 entries, we extracted 1035 rules having an IntImp value
above 0.99, out of the 2378 rules generated (with σs = 0.05 and σc = 0.60).
Finally, on the Housing data, containing n = 506 entries, 156 rules out of 263
were evaluated above 0.99 by IntImp, Apriori being run with σs = 0.02 and
σc = 0.55.
This phenomenon of loss of discriminant√power is illustrated in Figure 4,
in which we represent PI(1) = P (N (0, 1) < nr) for various values of n. This
figure shows the loss of discriminant power of the measure as n rises, although
r is not affected by such changes. For example, with n = 323, there are 991
rules evaluated above 0.999, 3540 when n is multiplied by 10, and 4205 when
n is multiplied by 100, out of the 5402 rules.
Using the third modeling does not solve the issue as can be seen in Figure 5.
In this situation almost all rules are evaluated above 0.95. On other rule sets,
as presented in Figure 6, the range of values that IntImp takes is wider.
In order to counter-balance this loss of discriminating power, [30] intro-
duce a contextual approach where ImpInd is centered and reduced on a case
database B, thus leading to the definition of the probabilistic discriminant
index (a monotonically increasing transformation of IntImp contextualized
on the data).
This index is defined as follows:
h i
PDI = P N (0, 1) > ImpIndCR/B

[13] propose an alternative solution by weighting IntImp through the use

of an inclusion index. This index is based on the entropy of experiments B/A
and A/B. We denote by H(X) = −px log2 px −px log2 px the entropy associated
with an event X. In [6] the most general form of the inclusion index is given
as:

 1
i(A ⊂ B) = (1 − H ∗ (B/A)α ) 1 − H ∗ (A/B)α 2α


where H ∗ (X) = H(X) if px > 0.5, H ∗ (X) = 1 otherwise. The α parameter

is chosen by the user. The value α = 2 is advised if one wants this index to be
tolerant to initial counter-examples, and we will use this value from now on.
Hence, [13] define the entropic intensity of implication as:
1
EII = [IntImp · i(A ⊂ B)] 2
 Generalized intensity of implication 429

Solarflare database
1.0
100 n
n 10 n
0.8
n/10

0.6
PI(1)

0.4

0.2

0.0
0 1000 2000 3000 4000 5000

Rule rank for Conf

Fig. 4. PI(1) values of the rules, in function of their rank for Conf on the So-
larflare database

Fig. 5. Values of IntImp in function of Fig. 6. Values of IntImp in function of

nab̄ for the Bcw database nab̄ for the Flag database
430 B. Vaillant et al.

The shift from H(X) to H ∗ (X) aims at discarding uninteresting situations,

such as pb/a < 0.5 or pa/b < 0.5, and complies with a predictive strategy. In a
targeting strategy, the value of pb/a should be compared to pb , and the value
of pa/b to pa .
The weighting of the implication of intensity by the inclusion index, al-
though effective, is problematic. The inclusion index is a measure of the dis-
tance to indetermination based on entropy, thus being null when pb/a = 0.5,
and so is EII. However, IntImp valuesq0.5 at independency. Hence EII is
2 2
not always null at independency: EII = 8 (1−H(A) 16)(1−H(B) )
if pa < 0.5 and
pb > 0.5, and is null otherwise.
Figures 7 to 9 show the effects of both approaches. Figure 8 shows the
difference in behavior of both anamorphosis of ImpInd, namely IntImp and
PDI. Since PDI is contextualized and takes into account the values of ImpInd
on the data, its distribution is smoother.

Fig. 7. Variations of EII Fig. 8. Variations of PDI Fig. 9. Variations of PDI

in function of IntImp in function of IntImp in function of EII

2.3 Adaptation of the entropic intensity of implication

We proposed two adaptations of EII in order to cope with the above mentioned
issues: Revised EII, denoted REII and Truncated EII, denoted TEII [25,26].
Our first proposal involves replacing IntImp by IntImp∗ in EII where:

IntImp∗ = max{2IntImp − 1; 0}

This will solve the previously highlighted problems, but has the drawback
of modifying the entire spectrum of values taken by EII:
1
REII = [IntImp∗ · i(A ⊂ B)] 2

Figures 10 and 11 show the joint distribution of EII and REII in function
of pab̄ . Three families of rules are presented, for the first and the last ones there
is no observable difference between the measures. On the contrary we see the
 Generalized intensity of implication 431

impact of the correction added in REII on the spectrum of values of EII for
the second family. In Figure 11, n is ten times smaller than in Figure 10. Here
for all three families, there are observable differences.
Our second proposal only nullifies the values of EII when pa pb̄ ≤ pab ≤
min{ p2a , p2b̄ }, without modifying its values otherwise. To achieve this, we in-
troduce Ht∗ (X) an adequate truncated version of H(X), and it a truncated
version of the inclusion index i. In order to take into account both predictive
and targeting strategies, a rule will have a non null evaluation by the inclusion
index, and hence by TEII when the following conditions are jointly met:

• pb/a > 0.5 (prediction) and pb/a > pb (targeting); i.e. pb/a > max(0.5, pb )
• pa/b > 0.5 (prediction) and pa/b > pa (targeting); i.e. pa/b > max(0.5, pa )
With these new conditions, TEII is null whenever the proportion of
counter-examples is above min pa pb̄ ; p2a ; p2b̄ :

1
TEII = [IntImp(A → B) × it (A ⊂ B)] 2
with:

 1
• it (A ⊂ B) = (1 − Ht∗ (B/A)α ) 1 − Ht∗ (A/B)α 2α ,


• Ht∗ (B/A) = H(B/A) if pb/a > max(0.5, pb ), Ht∗ (B/A) = 1 otherwise,

• Ht∗ (A/B) = H(A/B) if pa/b > max(0.5, pa ), Ht∗ (A/B) = 1 otherwise.

Fig. 10. Joint distributions of EII and REII in function of pab̄ , n = 2000
432 B. Vaillant et al.

Fig. 11. Joint distributions of EII and REII in function of pab̄ , n = 200

3 Statistical measures making reference

to indetermination
3.1 Probabilistic index
[3] propose IPEE , a probabilistic measure of deviation from indetermination
(or equilibrium). The authors implicitly use modeling 10 since they consider
CR Nab −0.5na
Nab ≡ B(na , 0.5) under an indetermination hypothesis, i.e. Nab = 0.5· √
na :

n − 0.5na
 
IPEE = P B(na , 0.5) > nab ≈ P N (0, 1) > ab √
 
0.5 · na
Under normal approximation, IPEE equals 0.5 at indetermination. This
measure corresponds to the probabilistic index associated with modeling 10
(see Table 1), where pb is replaced by 0.5. IPEE will hence inherit the weak
discriminating power of this kind of measure.
As shown in Figure 12, IPEE and EII may take significantly different
values for some rules.

3.2 Discriminant version

In order to enhance the discriminating power of IPEE, [5] proposed IP3E, in
which IPEE is weighted by the inclusion index, following a similar method as
used in EII:
0.5
IP3E = [IPEE × 0.5(1 + i(A ⊂ B))]
There is an important difference in the construction of these entropic mea-
sures. Indeed in IP3E, the inclusion index interacts with IPEE through the
 Generalized intensity of implication 433

Fig. 12. Variations on IPEE in function of IntImp

expression 0.5(1 + i(A ⊂ B)). This expression takes its values between 0.5
and 1, and equals 0.5 at indetermination. Hence, in this situation, the value
of IP3E is not nullified, as was the case for EII. As shown in Figure 13 the
contribution of this index is of less importance in this case. This can also be
seen in Figure 14 which compares IP3E to TEII.

4 Generalized statistical measures

Using the same approach as with descriptive measures [25], we generalize

statistical measures and evaluate the interestingness of a rule by comparing
its Conf to θ. This is done by considering in Table 1 that for each modeling
under H0 , the probability of an example, conditionally to na , is θ:

Nab ≡ B(na , θ)
The results of the thus adapted modelings 1 and 10 are immediate, and
those of modelings 2 and 3 are easily obtained through the use of the proba-
bility generating functions as detailed in [26] and recalled Table 1.
From these results, we propose a range of generalized measures (see
Table 1), which are constructed in the same way as described in section 2.
(i) CR
Generalized statistical measures are defined by GSI|θ = −Nab , while gener-
(i)
alized probabilistic measures are defined by GPI|θ = P (N (0, 1) > nCR
ab
):
434 B. Vaillant et al.

Fig. 13. Variations of IP3E in function of IPEE

Fig. 14. Comparison of the behavior of TEII and IP3E

 Generalized intensity of implication 435

• by establishing the law of Nab and Nab under the null hypothesis (H0 )
following the chosen modeling i, we can express a centered and reduced
(i)
index under H0 . This statistical index is denoted by GSI|θ .
• under standard conditions, the law of this index can be approximated to
the normal distribution, leading to the definition of a probabilistic mea-
sure, defined as the complement to 1 of the surprise of observing such
an exceptional value of the index under H0 . This probabilistic index is
(i)
denoted by GPI|θ .
(10 )
We will focus on two of these. The first one, GPI|θ , is associated with
modeling 10 and generalizes IPEE. For clarity reasons, it will be denoted
GIPE|θ (we here removed the last E, since the generalized measure no longer
makes reference to equilibrium). It corresponds to the chi-square goodness
of fit test, assessing whether or not the B/A distribution comes from the
(3)
distribution related to (θ; 1 − θ). The second one, GPI|θ , is associated with
modeling 3, and generalizes IntImp. It will thus be denoted by GIntImp|θ .
Using θ = 0.9 as lower reference value the generalized measures should
focus on rules having a confidence above this threshold. Clearly, we see in
Figures 15 and 16 that probabilistic indices stress the differences of evaluations
near this value, discarding rules far below to a null evaluation. On the contrary
rules above the reference tend to have a very good evaluation. Once more we
here see the importance of the use of a discriminant version of the probabilistic
measure.

(3) (3)
Fig. 15. Values taken by GSI|θ=0.9 in Fig. 16. Values taken by GPI|θ=0.9 in
function of Conf function of Conf
 436
Statistical and probabilistic indices summary

Modelings 1 and 10 Modeling 2 Modeling 3

1.1: na is fixed, Nab is 1’.1: pa is fixed, Nab is 2.1: Na ≡ B(n, pa ) 3.1: N ≡ P oi(n)
Principle randomized, and both randomized, and only 2.2: Nab ≡ B(na , pb ) |Na = na 3.2: Na ≡ B(n, pa ) |N = n
A and Ā are considered A is considered 3.3: Nab ≡ B(na , pb ) |N = n, Na = na
Law Nab under H0 H(n, na , pb ) B(na , pb ) B(n, na pb ) P oi(npa pb )
Law Nab under H0 H(n, na , pb ) B(na , pb ) B(n, na pb ) P oi(npa pb )

generalized counterparts
B. Vaillant et al.

Statistical index:
N −np p 0
CR ab a b Nab −npa pb Nab −npa pb Nab −npa pb
SI = −Nab SI(1) = − √np SI(1 ) = − √ SI(2) = − q SI(3) = −ImpInd = − √
a pa pb p b
npa pb pb npa pb (1−pa pb ) npa pb

Probabilistic index:
√ 0
PI = P (N (0, 1) > nCR
ab
) PI(1) = P (N (0, 1) < r n) PI(1 ) PI(2) PI(3) = IntImp = P (N (0, 1) > ImpInd)
PI = P (N (0, 1) < SI)

Generalized statistical and probabilistic indices summary

Modelings 1 and 10 Modeling 2 Modeling 3

1.1: na is fixed, Nab is 1’.1: pa is fixed, Nab is 2.1: Na ≡ B(n, pa ) 3.1: N ≡ P oi(n)
Principle randomized, and both randomized, and only 2.2: Nab ≡ B(na , θ) |Na = na 3.2: Na ≡ B(n, pa ) |N = n
A and Ā are considered A is considered 3.3: Nab ≡ B(na , θ) |N = n, Na = na
Law Nab under H0 H(n, na , θ) B(na , θ) B(n, na θ) P oi(npa θ)
Law Nab under H0 H(n, na , 1 − θ) B(na , 1 − θ) B(n, na (1 − θ)) P oi(npa (1 − θ))
Statistical index:
CR (1) N −npa (1−θ) (10 ) N
ab
−npa (1−θ) (2) Nab −npa (1−θ) (3)
GSI|θ = −Nab GSI|θ = − √ab GSI|θ =− √ GSI|θ = − √ GSI|θ = −GIndImp|θ
npa pa θ(1−θ) npa θ(1−θ) npa (1−θ)(1−pa (1−θ))
N
ab
−npa (1−θ)
=− √
npa (1−θ)

Probabilistic index:
0
(1) (1 ) (2) (3)
GPI|θ = P (N (0, 1) > nCR
ab
) GPI|θ GPI|θ = GIPE|θ GPI|θ GPI|θ = GIntImp|θ
= P (N (0, 1) > GIndImp|θ )
GPI|θ = P (N (0, 1) < GSI|θ )

Table 1. Modeling of the various statistical and probabilistic indices, and their
 Generalized intensity of implication 437

5 Discriminant versions of the generalized statistical

measures
The generalized statistical or probabilistic measures have, as the original ones
do, a weak discriminating power. In order to enhance these measures, we will
consider two approaches (cf. Section 2), one relying on weighting through the
use of an inclusion index, like [13], the other one being contextual, like [30].
In the first approach, we propose the more general expression of the en-
tropic generalized probabilistic index, EGPI|θ , which is defined as the product
of GPI|θ and an inclusion index. In order to remain coherent, we think it ad-
visable to define a generalized inclusion index gi|θ , using θ as the reference
value and not 0.5. This leads us to first define He |θ (X) a off-centered version
of the entropy H(X), being maximal when px = θ, and not when px = 0.5
(see Figure 18).

5.1 A off-centered version of the entropy

In order to define this off-centered version of the entropy, we propose the
following modification, so that the new index will take its maximal value 1
when px = θ.
This index is defined as follows:

e |θ (X) = −e
H px log2 pex − (1 − pex ) log2 (1 − pex )
where pex is:

px px + 1 − 2θ
pex = if px ≤ θ, pex = otherwise (see Figure 17)
2θ 2(1 − θ)
We call this He |θ (X) index the off-centered entropic index. It is clear that
H|θ (X) is not a strict entropy and that it must be seen as a penalization
e
function.
The behavior of this off-centered entropy index, illustrated in Figure 18
for a B/A distribution, leads to interesting perspectives in data mining.
It could, for example, be used in a tree induction process to assess the
quality of the prediction of the class variable conditionally to the predictive
variables, when such a class is boolean and has a very unbalanced distribu-
tion [22].
From the definition of this new index, we build H e |θ (B/A) and H e |θ (A/B)
as:
• to obtain H e |θ (B/A) from H(B/A), pb/a is replaced by peb/a defined as
follows:
pb/a pb/a + 1 − 2θ
peb/a = if pb/a ≤ θ, peb/a = otherwise
2θ 2(1 − θ)
438 B. Vaillant et al.

Fig. 17. Variation of pex in function of px

Fig. 18. Comparison of H(B/A) and H

e |θ=0.2 (B/A)
 Generalized intensity of implication 439

• to obtain H e |θ (A/B) from H(A/B), a first possibility involves replacing

pa/b by pea/b defined by:

pa/b pa/b + 1 − 2θ
pea/b = if pa/b ≤ θ, pea/b = otherwise
2θ 2(1 − θ)

This first possibility generalizes the inclusion index proposed in [13], which
can be retrieved using θ = 0.5.
• He |θ (A/B) could also be obtained from H(A/B), by using 1− pa ×(1−θ) as
pb
the reference, since pa/b = 1− ppa ×(1−pb/a ). In this case, when considering
b

independency (i.e. θ = pb ), the reference value for H e |θ (A/B) is pa . This

second possibility is the basis of a new version of the inclusion index.

5.2 Weighting approach

To define the entropic generalized probabilistic index, EGPI|θ , the generalized

e ∗ (B/A) and H
index of inclusion, gi|θ , is first defined. To this end, H e ∗ (A/B),
|θ |θ
are defined as:

e ∗ (X) = H
H e ∗ (X) = 1 otherwise
e |θ (X) if px > θ, H
|θ |θ

and gi|θ as:

h  1
i 2α
gi|θ = e ∗ (B/A)α
1−H e ∗ (A/B)α
1−H , with α = 2.
|θ |θ

From this, we deduce EGPI|θ which is a more discriminant version of

GPI|θ :
 1
EGPI|θ = GPI|θ × gi|θ 2
From this general expression of EGPI|θ , we can express particular instances
of generalized measures, such as EGINTIMP|θ , the entropic generalized
intensity of implication, and EGIPE|θ , the entropic generalized probabilis-
tic index of deviation:
• Modeling 3:
h i 12 1
(3) (3) 
EGINTIMP|θ = EGPI|θ = GPI|θ × gi|θ = GIntImp|θ × gi|θ 2

• Modeling 10 :
i 12 1
(10 ) (10 )
h 
EGIPE|θ = EGPI|θ = GPI|θ × gi|θ = GIPE|θ × gi|θ 2
440 B. Vaillant et al.
(3)
It must be noticed that both components of EGPI|θ = EGINTIMP|θ
refer to the same θ, which ensures the coherence of the measure. In particular,
(3) (3)
for θ = pb , GPI|pb = GIntImp|pb corresponds to IntImp and EGPI|pb =
EGINTIMP|pb is more coherent than EII.
(10 )
In the case θ = 0.5, GPI|θ = GIPEE|θ corresponds to IPEE and gi|θ
corresponds to i. It appears that EGIPE|0.5 is slightly different from IP3E =
1
[IPEE × 0.5 (i(A ⊂ B) + 1)] 2 , the entropic version of IPEE proposed by [3].
Their behavior, compared to their original counterparts, is represented in
Figure 19 (for n = 1000, pa = 0.05 and pb = 0.10). They were obtained
using 3 different values for θ, θ = pb = 0.1 (thus targeting independence),
θ = 2 pb = 0.2 (targeting situations in which B happens twice as often when A
is true) and θ = 0.5 (prediction).

Fig. 19. Behavior of the measures, as functions of pb/a for n = 1000, pa = 0.05 and
pb = 0.10

Figure 19 well illustrates how the θ parameter choice controls the behav-
ior of the measures. Furthermore, we can see the effectiveness of the para-
metrization of the statistical or probabilistic measures, making them more
discriminant.
In the specific case where the reference value is θ = pb , one could prefer
the second version of the inclusion index.
Figures 20 and 21 which compare both alternatives for the third modeling
to TEII put forward the best fitness of the second index.
 Generalized intensity of implication 441

Indeed in the first case all rules having pb ≥ pā /pb̄ ≥ 0.5 have a null
(3)
evaluation for EGPI|θ=pb .

(3) (3)
Fig. 20. Variations of EGPI|θ=pb in Fig. 21. Variations of EGPI|θ=pb in
function of TEII, first version of the en- function of TEII, second version of the
tropic coefficient entropic coefficient

0 0
If we consider now modeling 1 , and compare EGPI1|θ=0.5 to IP3E, we see
that its value is null under indetermination whereas IP3E still varies. The
range is therefore all the larger in our proposal (see Figure 22).

5.3 Contextual approach

In the contextual approach, GSI|θ is centered and reduced on a case database

B, and thus defines a contextual generalized probabilistic index, CGPI|θ :
CR/B
CGPI|θ = P (N (0, 1) > GSI|θ )
Depending on the modeling (1, 10 , 2 or 3), we can obtain different versions
of CGPI|θ :
(3)
• Modeling 3: GSI|θ corresponds to GIndImp|θ . Then CGPI|θ defines a
generalized probabilistic discriminant index, GPDI|θ :
(3) CR/B
GPDI|θ = CGPI|θ = P (N (0, 1) > GIndImp|θ )
It must be noted that GPDI|θ=pb = PDI
442 B. Vaillant et al.

(10 )
Fig. 22. Variations of EGPI|θ=0.5 in function of IP3E

(10 ) (10 )CR/B

• Modeling 10 : CGPI|θ = P (N (0, 1) > GPI|θ ) is a discriminant
version of GIPE|θ .
(10 )
Considering the case θ = 0.5, CGPI|θ=0.5 is an original contextual dis-
(10 )
criminant version of GPI|θ=0.5 = IPEE
(3)
Figure 23 compares the contextual and entropic version of GPI|θ=0.5
(3)
whereas Figure 24 compares the contextual approach using GPI|θ for the
independence and indetermination situations.

6 Conclusion and perspectives

Following modeling and coherence principles, we have proposed in previous
works an innovating framework from which a unified view of a large number
of statistical interestingness measures can be constructed, and which clarifies
some of the links between these measures.
This framework is the basis of the definition of new measures, namely
the generalized intensity of implication, generalized probabilistic discriminant
index, generalized probabilistic measure of deviation, and their entropic or
contextual discriminant counterparts, which all compare the confidence of a
 Generalized intensity of implication 443

Fig. 23. Comparison of the entropic and contextual approach (modeling 3), for
θ = 0.5

rule to a user defined reference parameter. We extended this concept and de-
fined an off-centered entropy. Its behavior within a supervised learning context
is currently under study, and should lead to new perspectives.
Based on a sound and comprehensive framework, this chapter illustrates
the use of parameterized measures within a data mining process. The behavior
of the parameterized measures is illustrated using classical datasets, and these
measures are compared to their original counter-parts. This study highlights
the different properties of each of them. Our proposal leads to the definition
of a set of measures, for which one may choose the most adapted one to user
needs and data specificities.

References
1. R. Agrawal, T. Imielinski, and A.N. Swami. Mining association rules between
sets of items in large databases. In P. Buneman and S. Jajodia, editors, ACM
SIGMOD International Conference on Management of Data, pages 207–216,
Washington, D.C., USA, 1993.
2. M. Bailleul and R. Gras. L’implication statistique entre variables modales.
Mathématiques, informatique et sciences humaines, (128):41–57, 1995.
3. J. Blanchard, F. Guillet, H. Briand, and R. Gras. Assessing the interestingness
of rules with a probabilistic measure of deviation from equilibrium. In J. Janssen
444 B. Vaillant et al.

Fig. 24. Comparison of the contextual approaches (modeling 3) making reference

at independence and indetermination

and P. Lenca, editors, The XIth International Symposium on Applied Stochastic

Models and Data Analysis, pages 191–200, Brest, France, 2005.
4. J. Blanchard, F. Guillet, H. Briand, and R. Gras. IPEE : Indice probabiliste
d’écart à l’équilibre pour l’évaluation de la qualité des règles. In Atelier Qualité
des Données et des Connaissances (EGC 2005), pages 26–34, Paris, France,
2005.
5. J. Blanchard, F. Guillet, H. Briand, and R. Gras. Une version discriminante de
l’indice probabiliste d’écart à l’équilibre pour mesurer la qualité des règles. In
R. Gras, F. Spagnolo, and J. David, editors, The Third International Conference
Implicative Statistic Analysis, pages 131–137, Palermo, Italy, 2005. Supplément
num. 15 de la Revue Quaderni di Ricerca in Didattica.
6. J. Blanchard, P. Kuntz, , F. Guillet, and R. Gras. Mesure de la qualité des
règles d’association par l’intensité d’implication entropique. Revue des Nouvelles
Technologies de l’Information (Mesures de Qualité pour la Fouille de Données),
(RNTI-E-1):33–43, 2004.
7. J. Blanchard, P. Kuntz, F. Guillet, and Gras R. Statistical Data Mining and
Knowledge Discovery, chapter Implication intensity: from the basic statistical
definition to the entropic version, pages 473–485. Chapman & Hall/CRC, 2003.
8. A. Freitas. On rule interestingness measures. Knowledge-Based Systems journal,
pages 309–315, 1999.
9. A. Freitas. Understanding the crucial differences between classification and
discovery of association rules - a position paper. In ACM SIGKDD Explorations,
volume 2, pages 65–69, 2000.
 Generalized intensity of implication 445

10. R. Gras. Contribution à l’étude expérimentale et à l’analyse de certaines ac-

quisitions cognitives et de certains objectifs didactiques en mathématiques. PhD
thesis, Université de Rennes I, 1979.
11. R. Gras. Panorama du développement de l’A.S.I. à travers des situations fon-
datrices. In R. Gras, F. Spagnolo, and J. David, editors, The third Interna-
tional Conference Implicative Statistic Analysis, pages 9–33, Palermo, Italia,
2005. Supplément num. 15 de la Revue Quaderni di Ricerca in Didattica.
12. R. Gras, R. Couturier, J. Blanchard, H. Briand, P. Kuntz, and P. Peter.
Quelques critères pour une mesure de qualité de règles d’association - un exem-
ple : l’intensité d’implication. Revue des Nouvelles Technologies de l’Information
(Mesures de Qualité pour la Fouille de Données), (RNTI-E-1):3–31, 2004.
13. R. Gras, P. Kuntz, R. Couturier, and F. Guillet. Une version entropique
de l’intensité d’implication pour les corpus volumineux. In H. Briand and
F. Guillet, editors, Extraction des connaissances et apprentissage (EGC 2001),
volume 1, pages 69–80. Hermes, 2001.
14. S. Guillaume. Traitement des données volumineuses, Mesures et algorithmes
d’extraction de règles d’association et règles ordinales. PhD thesis, Université
de Nantes, 2000.
15. R.J. Hilderman and H.J Hamilton. Knowledge discovery and interestingness
measures: A survey. Technical Report 99-4, Department of Computer Science,
University of Regina, october 1999.
16. R.J. Hilderman and H.J. Hamilton. Measuring the interestingness of discov-
ered knowledge: A principled approach. Intelligent Data Analysis, 7(4):347–382,
2003.
17. D. Janssens, G. Wets, T. Brijs, and K. Vanhoof. Adapting the CBA algorithm
by means of intensity of implication. Information Sciences, 173(4):305–318,
2005.
18. P. Kuntz, F. Guillet, R. Lehn, and H. Briand. A user-driven process for min-
ing association rules. In Principles of Data Mining and Knowledge Discovery,
volume 1910 of LNAI, pages 483–489. Springer, 2000.
19. J.B. Lagrange. Analyse implicative d’un ensemble de variables numériques;
application au traitement d’un questionnaire aux réponses modales ordonnées.
Revue de Statistique Appliquée, XLVI(1):71–93, 1998.
20. S. Lallich. Mesure et validation en extraction des connaissances à partir des
données. Habilitation à Diriger des Recherches – Université Lyon 2, 2002.
21. S. Lallich, P. Lenca, and B. Vaillant. Variations autour de l’intensité
d’implication. In R. Gras, F. Spagnolo, and J. David, editors, The Third In-
ternational Conference Implicative Statistic Analysis, pages 237–246, Palermo,
Italy, 2005. Supplément num. 15 de la Revue Quaderni di Ricerca in Didattica.
22. S. Lallich, P. Lenca, and B. Vaillant. Construction of an off-centered entropy for
supervised learning. In C. Skiadas, editor, The XIIth International Symposium
on Applied Stochastic Models and Data Analysis, Chania, Crete, Greece, 2007.
23. S. Lallich, E. Prudhomme, and O. Teytaud. Contrôle du risque multiple en
sélection de règles d’association significatives. In G. Hébrail, L. Lebart, and
J.-M. Petit, editors, Extraction et gestion des connaissances, volume 1-2, pages
305–316, Clermont-Ferrand, France, 2004. Cépaduès Editions.
24. S. Lallich and O. Teytaud. Évaluation et validation de l’intérêt des règles
d’association. Revue des Nouvelles Technologies de l’Information (Mesures de
Qualité pour la Fouille de Données), (RNTI-E-1):193–217, 2004.
446 B. Vaillant et al.

25. S. Lallich, B. Vaillant, and P. Lenca. Parametrised measures for the evaluation
of association rule interestingness. In J. Janssen and P. Lenca, editors, The
XIth International Symposium on Applied Stochastic Models and Data Analysis,
pages 220–229, Brest, France, 2005.
26. S. Lallich, B. Vaillant, and P. Lenca. A probabilistic framework towards the
parameterization of association rule interestingness measures. Methodology and
Computing in Applied Probability, 9(3):447–463, 2007.
27. P. Lenca, P. Meyer, B. Vaillant, and S. Lallich. On selecting interestingness
measures for association rules: user oriented description and multiple criteria
decision aid. European Journal of Operational Research, 184(2):610–626, 2008.
28. P. Lenca, P. Meyer, B. Vaillant, P. Picouet, and S. Lallich. Évaluation et analyse
multicritère des mesures de qualité des règles d’association. Revue des Nouvelles
Technologies de l’Information (Mesures de Qualité pour la Fouille de Données),
(RNTI-E-1):219–246, 2004.
29. P. Lenca, B. Vaillant, P. Meyer, and S. Lallich. Quality Measures in Data
Mining, volume 43 of Studies in Computational Intelligence, Guillet, F. and
Hamilton, H.J., Eds., chapter Association rule interestingness measures: exper-
imental and theoretical studies, pages 51–76. Springer-Verlag Berlin Heidelberg,
2007.
30. I.C. Lerman and J. Azé. Une mesure probabiliste contextuelle discriminante de
qualité des règles d’association. In M.-S. Hacid, Y. Kodratoff, and D. Boulanger,
editors, Extraction et gestion des connaissances, volume 17 of RSTI-RIA, pages
247–262. Lavoisier, 2003.
31. I.C. Lerman, R. Gras, and H. Rostam. Elaboration d’un indice d’implication
pour les données binaires, i et ii. Mathématiques et Sciences Humaines, (74,
75):5–35, 5–47, 1981.
32. K. McGarry. A survey of interestingness measures for knowledge discovery.
Knowledge Engineering Review Journal, 20(1):39–61, 2005.
33. G. Piatetsky-Shapiro. Discovery, analysis and presentation of strong rules. In
G. Piatetsky-Shapiro and W.J. Frawley, editors, Knowledge Discovery in Data-
bases, pages 229–248. AAAI/MIT Press, 1991.
34. G. Ritschard. De l’usage de la statistique implicative dans les arbres de classifi-
cation. In R. Gras, F. Spagnolo, and J. David, editors, The third International
Conference Implicative Statistic Analysis, pages 305–315, Palermo, Italia, 2005.
Supplément num. 15 de la Revue Quaderni di Ricerca in Didattica.
35. G. Ritschard and D.A. Zighed. Implication strength of classification rules. In
F. Esposito, Z.W. Ras, D. Malerba, and G. Semeraro, editors, 16th International
Symposium on Methodologies for Intelligent Systems, volume 4203 of LNAI,
pages 463–472, Bari, Italy, 2006. Springer.
36. E. Suzuki. In pursuit of interesting patterns with undirected discovery of ex-
ception rules. In S. Arikawa and A. Shinohara, editors, Progresses in Discovery
Science, volume 2281 of Lecture Notes in Computer Science, pages 504–517.
Springer-Verlag, 2002.
37. E. Suzuki and Y. Kodratoff. Discovery of surprising exception rules based on
intensity of implication. In J. M. Zytkow and M. Quafafou, editors, Principles
of Data Mining and Knowledge Discovery, volume 1510 of Lecture Notes in
Artificial Intelligence, pages 10–18, Nantes, France, September 1998. Springer-
Verlag.
38. P-N. Tan, V. Kumar, and J. Srivastava. Selecting the right objective measure
for association analysis. Information Systems, 4(29):293–313, 2004.
 Generalized intensity of implication 447

39. B. Vaillant. Mesurer la qualité des règles d’associations : études formelles et

expérimentales. PhD thesis, ENST Bretagne, Université de Bretagne Sud, 2006.
40. B. Vaillant, P. Lenca, and S. Lallich. A clustering of interestingness measures.
In E. Suzuki and S. Arikawa, editors, Discovery Science, volume 3245 of Lecture
Notes in Artificial Intelligence, pages 290–297, Padova, Italy, 2004. Springer-
Verlag.
The TVpercent principle
for the counterexamples statistic

Ricco Rakotomalala1 and Alain Morineau2

1
Eric Laboratory – Bron, France
[email protected]
2
Modulad – Rocquencourt, France
[email protected]

Summary. Our aim is to put into practice the principle of test value percent crite-
rion to the counterexamples statistic, which is the basis of the well-known statistical
implicative analysis approach. We show how to compute the test value in this con-
text; what is the connection with the intensity of implication measure, on the one
hand; and the index of implication, on the other hand. We evaluate the behavior
of these measures on a large dataset comprising several hundred of thousands of
transactions. We evaluate especially the discriminating capacity of the measures, in
relation to specialized measure such as the entropic intensity of implication.

Key words: Association rule, Measure, TVpercent, Intensity of implication.

1 Introduction
Since the work of Agrawal and Srikant (1994) [1], the association rule mining
has received a great deal of attention and is became one of the most popular
method in the knowledge discovery community. This approach allows to pro-
duce implication rules such as “If A Then C”, where A and C are sets of items
or products in the analysis of market basket data. The meaning of the rule is
“whenever a set of transactions contains A, than it probably contains also C”.
Even if the association rule mining is very powerful, there is a pitfall which
can call into question its use: the number of generated rules can be very high,
it becomes difficult to distinguish the most interesting rules [13]. In this con-
text, it is important to have a numerical indicator which makes it possible to
propose the most relevant rules quickly, but also to validate them, so as to keep
only the rules which show a real causation. There are many proposals of rule
quality measurements these last years. Among them, we are interested in the
intensity of implication measure based on the counterexamples statistic [8, 9].
In order to transform the regularity, the concomitant occurrence of the
itemsets, into causation, the implication rule, we count the counterexamples

R. Rakotomalala and A. Morineau: The TVpercent principle for the counterexamples statistic,
Studies in Computational Intelligence (SCI) 127, 449–462 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
450 Ricco Rakotomalala and Alain Morineau

of the rule. The rule “If A Then C” is all the more relevant as it has few
counterexamples. The intensity of implication is a measurement that is based
on this idea. It was implemented in many domains. It is available on several
bench test softwares [14, 23].
The intensity of implication measure relies on a classical hypothesis-testing
scheme. The idea is not to test the absence or the presence of a real link be-
tween A and C, but rather to what extent we deviate from the reference
situation described by the null hypothesis. In this context, we often compute
the p-value of the test. It shares the property of any classical statistical index
in data mining context: for the same constant proportion, the value inop-
portunely increases with the number of observations. In certain situations,
when the number of examples is very high, the p-value cannot be computed
correctly because we exceed the accuracy of the common statistical libraries.
Thus, all the rules correspond to the maximum value of measurement. It is
not possible to distinguish the relevant rules.
In this paper, we put into practice the test-value percent principle on the
counterexamples statistic. The test-value is the expression of the p-value into
the number of standard deviation of the Gaussian distribution [15, 17]. In
a recent paper, we proposed a normalized version of the Test-Value named
TVpercent criterion [18]. Applications to association rules proved it to be
an interesting criterion to eliminate statistically uninteresting rules without
being influenced by the number of occurrences. This criterion remains com-
prehensible and discriminating even if we treat a huge database. It suggests a
threshold that allows to eliminate irrelevant rules. It enables to compare rules
from different databases.
The organization of this paper is as follows. In section 2, we recall the
computation formulas for the intensity of implication which is associated to
the counterexamples statistic. In the section 3, we quickly show the TVper-
cent framework. Then, we extrapolate the calculation of the TVpercent to the
counterexamples statistic. A new measure, ve , is described. In section 4, we
study the behavior of this measure on a large size database (340183 transac-
tions and 468 items). We compare the TVpercent to the standard intensity
of implication and the entropic intensity of implication which is designed for
rules with high supports [2]. We conclude in the section 5.

2 Intensity of implication and counterexamples statistic

2.1 Computing the intensity of implication

The intensity of implication is based on the counterexamples statistic. In a rule

of the type “If A then C”, we check if the number of observed counterexamples
is significantly infrequent. We use a statistical hypothesis testing framework.
We must then define [22]: a parameter which we estimate; a formulation of
the null hypothesis, the reference situation; the statistical distribution of the
 The TVpercent principle for the counterexamples statistic 451

Rule A Ā Total
C
C̄ nac̄ nc̄
Total na n
Table 1. Contingency table — Number of transactions for a rule “If A then C”

parameter estimate under the null hypothesis; and then, an indicator which
measures the deviation of the observed data from the null hypothesis.
In the Gras’ approach [7], the statistical parameter is the number of coun-
terexamples Nac̄ to a rule. Its estimation is naturally the number of observed
counterexamples nac̄ (Table 1). The null hypothesis is the independence be-
tween the antecedent and the consequent of the rule. In this situation, the
probability πac̄ , i.e. e probability for obtaining counterexamples if the an-
tecedent of the rule is true, is equal to πa × πc̄ . It is the product of two
marginal probabilities, they can be estimated by nna × nnc̄ . The expectation
of the random variable Nac̄ under the null hypothesis is Λ = n × πa × πc̄ ,
estimated by λ = n × nna × nnc̄ .
Although we have a hypothesis testing framework, the goal is not to ac-
cept or reject the null hypothesis, but the characterization of the deviation
from this reference. In our situation, we want to characterize in what extent
the observed number of counterexamples nac̄ is less than λ. Various modeling
approaches are available. We can use a hypergeometric, binomial or Poisson
distribution. We mainly study here the third model with the Poisson distrib-
ution [16]. More than a simple approximation of the other distributions, this
sampling scheme is interesting because it treats in a non symmetrical way the
positive and negative associations, which enables to show a causation.
The p-value, the probability of obtaining a result at least as extreme or
impressive as that obtained for the data, assuming the null hypothesis is true,
is computed with a Poisson distribution with the parameter λ. The critical
region for Nac̄ is defined as the interval [0, nac̄ ]. The intensity of implication,
Ie , is the complement to 1 of the p-value of the test:
nac̄
X λm −λ
Ie = 1 − e (1)
m=0
m!
Numerical example We use an example described in the Ritschard’s pa-
per [20]. We want to characterize a rule where n = 273, na = 76, nc̄ = 153,
nac̄ = 28 (Table 2), and λ = 42.59, the intensity of implication is Ie = 0.9884.

2.2 Practical computation of the intensity of implication

When λ is large, about 18, it is possible to use the Gaussian approximation

of the Poisson distribution, the mean and the variance are both λ. We then
452 Ricco Rakotomalala and Alain Morineau

Rule A Ā Total
C
C̄ 28 153
Total 76 273
Table 2. Example of a contingency table from Ritschard [20]

compute the index of implication, which is the standardized value of the ob-
served number of counterexamples: we use a continuity correction factor.
nac̄ + 0.5 − λ
ie = √ (2)
λ
The approximation of the intensity of implication using the Gaussian CDF
(cumulative distribution function) Φ is defined as follows:

Ia = 1 − Φ[ie ] (3)
Numerical example We take again the above example (Table 2). The index
of implication is ie = 28+0.5−12.59
√
42.59
= −2.16. The approximated intensity of
implication is Ia = 0.9846. We note that the true intensity (Ie ) and the
approximated intensity (Ia ) are similar. Very often, only the approximate
formulation is referred in publications.
There is a drawback for the utilization of the intensity of implication Ia if
the support of the rule nac is large: the computed value of Ia is mechanically
equal to 1 because the available libraries to compute the Gaussian cumula-
tive distribution function (CDF) are not enough accurate. For instance, if the
index of implication ie is less than −6.2, the Excel© spreadsheet gives system-
atically an intensity of implication equal to 1. We tested several libraries [11],
none could not significantly exceed this limitation.
Using the original formulation —Ie , Poisson distribution, equation (1)—
can slightly improve the results. But we deal with very small values of which
are badly handled by computation libraries. In our example (Table 1), if we
multiply all the values in the table by 4, counts would be equal to n = 1092 and
nac̄ = 112. These values are not excessive. However, we find with the exact for-
mula Ie = 0.99999879, and with the approximate formula Ia = 0.999995367.
It becomes very difficult to distinguish the interesting rules. This problem is
designed by the loss of the discriminating power of measurements. Elegant
solutions were suggested, especially with the concept of the entropic intensity
of implication where we balance the intensity of implication with an index of
inclusion [10]. We present this approach below in the experimentation section.
 The TVpercent principle for the counterexamples statistic 453

Fig. 1. From the p-value to the test value

3 TVpercent criterion based on the counter-examples

statistic

3.1 Test value and normalized test value

Test value The test value also relies on a statistical framework. We try to
compare the observed parameter with the theoretical parameter under the
null hypothesis, which is the independence between the antecedent and the
consequent of a rule.
We mainly use the p-value p to characterize the strength of the deviation
between the observed number of counterexamples and the theoretical num-
ber of counterexamples under the reference situation. The p-value can take
very small values, close to zero, thus not very comprehensible as soon as one
deviates from the situation of reference with a large database size. In order
to obtain a better adapted measure, easily interpretable, we replace it by
the number of standard deviations of the standardized Gaussian distribution
which should be exceeded to cover the computed p-value (Figure 1). We call
the test value this criterion (equation 4).

T V = Φ−1 (1 − p) (4)
This criterion is often used to compare proportions or conditional average
for the characterization of the clusters built with clustering process [15].

Numerical example In our example (Table 2), the computed p-value for
nac̄ = 28 and λ = 42.59 with the Poisson distribution is p = 1 − Ie = 0.0116,
the corresponding test value is T V = 2.2701. This value is comparable to the
index of implication of which the negative, ve , can be considered as a rough
approximation of the test value ve = −ie = 2.16.

TVpercent — A normalized test value In the context of the knowl-

edge discovery, we handle a very large size databases, larger than the usual
sample size in statistical inference. We must treat two kinds of problems: as
we have noted it, we obtain a very small p-value, hardly handled by the al-
gorithms of the standard libraries for CDF computation; more disturbing is
454 Ricco Rakotomalala and Alain Morineau

a well known phenomena by the statisticians, i.e. when the sample size in-
creases, a small deviation from the values of the parameter under the null
hypothesis becomes significant, even if it corresponds to a statistical artifact.

Numerical example For instance, we assume that the number of coun-

terexamples is equal to nac̄ = 40. The computed p-value is 0.37420 (Ia =
1 − 0.374 = 0.62580). The rule seems not to be relevant. When we multiply
all the values of the table by 10, we obtain the p = 0.10890; if we multiply
by 100, the p-value becomes p = 0.00004. It seems now that the rule is very
relevant and the association is strong.

In order to avoid this pitfall, we have proposed a normalized test value [18].
The measure becomes independent of the real size of the database. We do not
forget that the main goal of the measure is to rank the rules in decreasing
relevance, and secondarily to suggest a cut value below which we can consider
that the rule does not bring relevant information.
The main idea is to set a priori the size of the dataset to 100. This value
corresponds to a reasonable size of the samples used when statistical inference
and hypothesis testing were historically developed. The value 100 is surely
an arbitrary value. But it is not more arbitrary than the usual confidence
level used in statistical inference (e.g. 5%, 1%, etc.). These confidence levels
are the results of the experiments of Fisher [6]. Indeed, in a not well-known
process depicted by Poitevineau [19], Fisher had hesitated for setting the
right value of the confidence level [5, 6]. In effect, the appropriate value of the
confidence level relies on the studied problem, the goal of the statistician, and
the characteristics of the dataset, especially the dataset size. From this point
of view, a criterion which allows to sorting the rules is surely essential. Using
the same criterion in order to mechanically accept or reject a rule is doubtful.
The original process to compute the normalized test value is a Monte Carlo
sampling approach. We draw randomly with replacement 100 examples from
the database and we compute the p-value p = 1 − Ie from equation (1) for the
corresponding 2 × 2 contingency table (Poisson approximation). We repeat
this process and then compute the average of the p-values p̄. In the last step,
the normalized test value is computed from this average T Vnorm = Φ−1 (1− p̄).
If we use a sufficient number of repetition (e.g. 2000 samples of 100 examples
with replacement), we obtain a stabilized value of the test value.
This process is known also as the bootstrap procedure, but the size of the
sample is arbitrarily set to 100 here, and not equal to the dataset sample size.
This criterion makes it possible to rank the rules computed on a database.
It has also the advantage, since it proceeds to the evaluation of the rules
in an unique reference (100 examples), of allowing the comparison of rules
computed on several similar databases, for example, on databases of different
size extracted on successive dates.
Practical computation of the TVpercent
 The TVpercent principle for the counterexamples statistic 455

Fig. 2. Barycentric approximation of the TVpercent

Règle A Ā Total
C
C̄ 10.25 56.04
Total 27.83 100
Table 3. Table brought back to 100 of table 2

The Monte Carlo migh be a good approach in order to obtain a stabilized

value of the criterion, but it is time consuming, especially because it needs a
repeated access to the database. We thus must use an approximation which is
quickly computed, it must rank the rules in the same order than the original
approach.
We use an interpolation procedure in order to compute the TVpercent.
In the first step, we bring back the values of the contingency table to 100
(e.g. from table 2 to table 3). In the second step, we compute the p-value
for each integer value which surrounds the decimal values in the table (e.g.
10.25 is surrounded by 10 and 11). Then we compute the average of 8 p-values
obtained for the 8 corners of the cube (Figure 2), which correspond to the 8
contingency tables computed from the integer values. From this average p100
of p-values, we compute the test value which is the TVpercent, equation (5).

T Vprecent = Φ−1 (1 − p100 ) (5)

The approximation can appear naive. But it is sufficient for our goal which
is to compare the relevance of the rules and to rank them [18]. We can use
a more accurate procedure, but it should not to be at the expense of the
computing time which is a major constraint in our context.

3.2 TVpercent criterion on the counter-examples statistic

In our first work, we use the TVpercent criterion for the co-occurrence of the
antecedent and the consequent (nac ). We used a hypergeometric distribution
456 Ricco Rakotomalala and Alain Morineau

nac̄ na nc̄ p − value

10 27 56 0.1127
10 27 57 0.1007
10 28 56 0.0890
10 28 57 0.0788
11 27 56 0.1769
11 27 57 0.1602
11 28 56 0.1437
11 28 57 0.1290
Table 4. The 8 configurations to evaluate for the barycentric estimation

[18]. In fact, the TVpercent principle can be extented to other parameters,

such as the counterexamples statistic and the Poisson distribution.
The detail of the process is the following. We bring back the original con-
tingency table (e.g. Table 2) to n = 100 (e.g. Table 3). We enumerate the 8
configurations which surrounds the decimal values. For each configuration, we
compute the p-value using the Poisson distribution (e.g. Table 4). We compute
the average of these p-values using the barycentric approximation (Figure 2).
Then, the test value is computed from the Gaussian CDF.

Numerical example From Ritschard’s example (Table 2), the 8 configura-

tions to evaluate are displayed in the table 4. The average of the p-values
is p100 = 0.1410 and the corresponding test value is T V percent = Φ−1 (1 −
0.1410) = 1.0759.

3.3 Test value and index of implication

When we handle large database, Ia , using the Gaussian CDF in order to get an
approximation of the intensity of implication, is widely used. In our context,
is the index of implication ie is a good approximation of the test value?
For a good understanding of the problem, we must not forget that the test
value T V percent is computed from the p-value p100 with a Gaussian CDF
(symmetrical distribution), the p-value is computed from the statistical for-
mulation with a Poisson distribution (a priori non symmetrical distribution).
The approximation between the test value and the standardized counterex-
amples statistic is satisfactory if these two distributions are symmetrical. The
Poisson distribution becomes approximately symmetrical when we have a high
value of the λ parameter i.e. we have a large database. In our context, because
we bring back n to 100, the approximation is not accurate.

Numerical example In the Table 3, λ = 27.83×56.04

100 = 15.6. The negative
index of implication is ve = −ie = − 10.25+0.5−15.6
√ = 1.2267. The deviation
(15.6)
from the TVpercent (1.0759) is considerable.
 The TVpercent principle for the counterexamples statistic 457

4 Experiments

4.1 Description of the experimentation

Database In order to evaluate the behavior of the various measures, we use

a large size database (ACCIDENT) which counts accident locations [12]. The
number of transactions is 340183, and the number of items is 468. Our goal is
to study the behavior of the T V percent criterion in relation to state of the art
measures such as intensity of implication and entropic intensity of implication.
Intensity of implication and entropic intensity of implication We
want to compare the TVpercent criterion with the intensity of implication Ia
measure in a large database context. It seems also interesting to compare our
measure to the version of the intensity of implication dedicated to the large
databases, it is the entropic intensity of implication [2, 10]. It is defined by
p
IEa = Ia × h (6)
where h, the index of inclusion, is equal to
r
nac nāc̄
h = (1 − H( )) × (1 − H( ))
na nc̄
H(x) is a modified Shannon entropy function [10]
• if x < 0.5, H(x) = 1 + 0.5 × [x × log2 (x) + (1 − x) × log2 (1 − x)];
• if x ≥ 0.5, H(x) = −0.5 × [x × log2 (x) + (1 − x) × log2 (1 − x)]

Computing the rules and the measures We use the Borgelt’s imple-
mentation for the rule generation [3]. It is available on the web site of the
author3 . Its implementation is very efficient but it computes rules with one
item only in the consequent. The parameters of the software are classical, we
can choose the minimum support, the minimum confidence and the maximum
length (number of items) of the rules.
The rules are then post processed in the Excel© spreadsheet. We compute
the various measures that we want to evaluate in this paper (Ie , Ia , IEa ,
T V percent, ie ). In spite of some doubt about the accuracy of this spreadsheet,
our background shows that the available very competitive specialized libraries4
are not really more accurate. This spreadsheet is considered to be adequate
in our exploratory study.
Evaluation framework Our aim is to check the concordances and the
discordances between our measure (T V percent) and the state of the art mea-
sures [10]. The first approach is to check if the various measures rank the rules
in the same way. A scatter plot allows to check that. We can also compute

3
http://fuzzy.cs.uni-magdeburg.de/~borgelt/apriori.html
4
e.g. STATLIB library — http://lib.stat.cmu.edu/index.php
458 Ricco Rakotomalala and Alain Morineau

Indicator Value
Number of transactions 340183
Number of items 468
Maximum length of rules 3
Support minimum 10%
Confidence minimum 75%
Number of rules 17212
Table 5. Characteristics of our experiments on the ACCIDENT dataset

a numerical indicator such as the Spearman’s rank correlation coefficient [21]

(correlation computed on ranks). But, if it is easy to calculate, we must use
it with caution, a numerical indicator often hides a complex situation such as
a non linear relation, etc. We will use it only as a rough indication.
Because in real studies, the rules are validated by experts. It is important
to carefully study how the first rules are ranked by the measure. An expert,
even if he is really persevering cannot check a hundred rules. In this paper,
we focus on the first 20 rules. This kind of evaluation is similar to ROC curve
analysis where, in certain circumstances, the AUCn criterion, focused on the
first n examples, is more relevant [4].

4.2 Results and comments

The characteristics of the database and the computed rule set are described
in the Table 5. The parameters of the algorithm have been chosen after sev-
eral attempts. We note that the results of the various attempts are not in
contradiction with the results presented here.
With a minimum support of 10%, the support of the rules runs from
34018 to 340183 transactions. The number of counterexamples nac̄ of a rule
runs from 0 (no counterexample) to 81742 (the support of the rule is 258408
in this situation). In this context, the accuracy of the computation is very
important for the used measures.
The exact intensity of implication Ie : The exact formulation of the
intensity of implication, equation (1), can be computed only on 2864 rules
(among 17212 rules). The Excel© implementation of the Poisson CDF cannot
handle some values. Even if we use some tricks to improve the accuracy, we
doubt we can really improve the exact formulation on a large database.
The approximated intensity of implication Ia : The approximate for-
mulation using the Gaussian CDF, equation (3), is more robust. It can be
computed on all the rules. Another drawback appears. When the index of
implication is very small (in some situations it can be equal to −209.9), the
Gaussian CDF implemented in the spreadsheet is mechanically equal to 0. So
the approximate intensity of implication is 1 for 9960 rules among 12712.
At the beginning, we have thought that there was a specific problem of the
spreadsheet. But we found the same limitations with the specialized libraries
 The TVpercent principle for the counterexamples statistic 459

Fig. 3. Scatterplot of the TVpercent and the entropic intensity of implication

implemented in very powerful languages for numerical calculation such as

FORTRAN5 . For ie ≤ −6.2, the p-value cannot really be computed from the
index implication with a Gaussian CDF. Using a measure which is derived
from Ia for detecting the relevant rules is not a powerful approach in this
context.
The entropic intensity of implication IEa : The entropic intensity of
implication, equation (6), allows to significantly overcome this drawback. The
index of inclusion h takes over, to some extent, the intensity of implication
when this one is saturated. So, in our experiments, the number of rules where
IEa = 1 is 79 (among 12712). The discriminating power of the measure is
maintained. But there is some problem nevertheless. Indeed, if we rely on a
manual expertise, the presentation of the first 20 rules will depend primarily
on the sorting algorithm and not on the sorting criterion, this is not very
satisfactory.
TVpercent for the counterexamples statistic: The computation of
the TVpercent is theoretically slower than the other measures since we com-
pute 9 CDF (8 Poisson distribution and 1 Gaussian distribution). But it is
not really perceptible on our set of rules. None capacity overflow have been
observed. The TVpercent has a range between −3.12 to 4.58. The rules which
have equal TVpercent are those of which we observe the same values (na , nc̄ ,
nac̄ ). When we compare the TVpercent measure with the other measures, we
note that the correlation with the entropic intensity of implication is small
(0.21). There is little concordance between these measures (Figure 3).

5
e.g. STATLIB library (http://lib.stat.cmu.edu/index.php). The available imple-
mentations are described in a book [11].
460 Ricco Rakotomalala and Alain Morineau

Fig. 4. Scatterplot of Index of implication and TVpercent

When we deeply studied the results, we found that we have not a symmet-
rical situation. On the 20 first rules according to the TVpercent, 12 have the
maximal value of entropic intensity of implication IEa = 1. At the opposite,
there are 79 rules with IEa = 1, the best rules according to the TVpercent
are hidden among these rules.
TVpercent and the normalized index of implication We had al-
ready noted above the similarity between the index of implication and the
TVpercent, we had also noticed that the approximation was not very precise
on small dataset. What is the case when we bring back the values to n = 100.
Nevertheless, we calculated the negative of the index of implication directly
on the sample size brought back to 100. Indeed, even if the approximation of
the TVpercent with the index of implication is bad, perhaps they rank the
rules in a similar way?
The scatterplot shows that there is little discordance between the two mea-
sures, even if the approximation is not really accurate. The relation between
these measure is clearly non linear but monotonic (Figure 4). This visual im-
pression is corroborated with the correlation computed between the rank of
the two measures (the Spearman rank correlation) which is equal to 0.999.
When we focus on the 20 first rules according to the normalized index of impli-
cation, we have found the 18 best rules according to the TVpercent criterion.
In a real situation where we present the rules to a human expert, these two
criteria will approximately propose the same rules.
The principal difference between these two indicators is in the determina-
tion of the statistical valid rules. If we use critical values associated to usual
significance levels (e.g. 2.32 for a significance level of 1%, etc.), because the
TVpercent is always larger, we keep more rules with the TVpercent than the
normalized index of implication. It is not really a disadvantage of the nor-
malized index of implication. We have seen that using an arbitrary threshold
value in order to keep or remove rules must be made with caution. If the
user wants nevertheless to use this procedure, he must take into account this
 The TVpercent principle for the counterexamples statistic 461

difference when he specifies the thresholds in order to avoid the elimination

of interesting rules.

5 Conclusion
In this paper, we have generalized the TVpercent criterion to the counterex-
amples statistic. The main improvement of this new measure is the handling of
the rules computed on large databases. We preserve the discrimination power
of the measure i.e. the ability to rank rules without ties according to the mea-
sure. We can rank a great number of rules. In this way, it extends the field
of application of the intensity of implication and constitutes an alternative to
the entropic intensity of implication.
The second main result of this work is the similarity between the TVper-
cent and the normalized index of implication. Although the index underesti-
mates the true value of the TVpercent, it ranks the rules in same way, and
most of all, it points up the same rules if we are interested in the best rules
according to these criteria.
Of course, these conclusions rely mainly on experimental evaluation. We
studied the behavior of these measures on various parameters settings of the
rule extraction algorithm, corresponding to the post processing of more or
less large number of rules. The results described above were not called into
question. However, it would be interesting to complete this study on other
databases with different characteristics, for example with few transactions
and a very large number of items.

References
1. R. Agrawal and R. Srikant. Fast algorithms for mining association rules. In
Proceedings of the 20th VLDB Conference, pages 487–499, 1994.
2. J. Blanchard, P. Kuntz, F. Guillet, and R. Gras. Mesures de qualité pour
la fouille de données, volume E-1, chapter Mesure de qualité des règles
d’association par l’intensité d’implication entropique. Revue des Nouvelles Tech-
nologies de l’Information, 2004.
3. C. Borgelt and R. Kruse. Induction of association rules: A priori implementation.
In 15th Conference on Computational Statistics, 2002.
4. T. Fawcett. Roc graphs: Notes and practical considerations for researchers.
Technical Report - HP Laboratories, 2003.
5. R.A. Fisher. Statistical Methods, Experimental Design, and Scientific Inference,
chapter The design of experiments. Oxford University Press, 1990. 1st edition,
1935, London, Oliver and Boyd.
6. R.A. Fisher. Statistical Methods for Research Workers. Oxford University Press,
14 edition, 1990. 1st edition, 1925, London, Oliver and Boyd.
7. R. Gras. Contribution à l’étude expérimentale et à l’analyse de certaines ac-
quisitions cognitives et de certains objets didactiques en mathématiques. Thèse
d’Etat, 1979.
462 Ricco Rakotomalala and Alain Morineau

8. R. Gras. L’implication statistique - Nouvelle méthode exploratoire des données.

La pensée sauvage, 1996.
9. R. Gras. La fouille dans les données par la méthode d’analyse implicative, chap-
ter Les fondements de l’analyse statistique implicative. Ecole polytechnique de
l’Université de Nantes, IRIN et IUFM Caen, 2004.
10. R. Gras, P. Kuntz, R. Couturier, and F. Guillet. Une version entropique de
l’intensité d’implication pour les corpus volumineux. In Actes de la Conférence
Extraction et Gestion des Connaissances, EGC’2001, pages 69–80. Revue Ex-
traction et Gestion de la Connaissance, 2001.
11. P. Griffiths and I.D. Hill. Applied Statistics Algorithms. Ellis Horwood:
Chichester, 1985.
12. K. Guerts, G. Wets, T. Brijs, and K. Vanhoof. Profiling high frequency accident
locations using association rules. In Proceedings of the 82nd Annual Transporta-
tion Research Board, 2003.
13. F. Guillet, editor. Mesures de la qualité des connaissances en ECD. EGC’2004,
RNTI, 2004.
14. X-H. Huynh, F. Guillet, and H. Briand. Une plate-forme exploratoire pour la
qualité des règles d’association : apports pour l’analyse implicative. In Actes
des Troisièmes Rencontres de l’Analyse Statistique Implicative, pages 339–349,
2005.
15. L. Lebart, A. Morineau, and M. Piron. Statistique Exploratoire Multidimension-
nelle. Dunod, Paris, 1995.
16. I. Lerman, R. Gras, and H. Rostam. Elaboration et évaluation d’un indice
d’implicatioon pour données binaires. Mathématique et Sciences Humaines,
(74):5–35, 1981.
17. A. Morineau. Note sur la caractérisation statistique d’une classe et les valeurs-
tests. Bull. techn. du Centre de Statis. et d’Infor. Appl., (2):20–27, 1984.
18. A. Morineau and R. Rakotomalala. Critère vt100 de sélection des règles
d’association. In Actes de Extraction et Gestion de Connaissances, EGC’2006,
pages 581–592, 2006.
19. Jacques Poitevineau. L’usage des tests statistiques par les chercheurs en psy-
chologie : aspects normatif, descriptif et prescriptif. Mathématiques et Sciences
Humaines, 3(167):5–25, 2004.
20. G. Ritschard. De l’usage de la statistique implicative dans les arbres de classifi-
cation. In Troisième Rencontre Internationale - Analyse Statistique Implicative,
pages 305–316, 2005.
21. G. Saporta. Probabilités, Analyse de Données et Statistique. Technip, 1990.
22. B. Vaillant, S. Lallich, and P. Lenca. Modeling the counter-examples and as-
sociation rules interestingness measures behavior. In S.F. Crone, S. Lessman,
and R. Stahlbock, editors, The 2006 International Conference on Data Mining,
pages 132–137, 2006.
23. B. Vaillant, P. Picouet, and P. Lenca. An extensible platform for rule quality
measure benchmarking. In Human Centered Processes (HCP’03), Distributed
Decision Making and Man Machine Cooperation, pages 187–191, 2003.
User-System Interaction for Redundancy-Free
Knowledge Discovery in Data

Rémi Lehn, Henri Briand, and Fabrice Guillet

Laboratoire d’Informatique de Nantes Atlantique

Equipe COnnaissances & Décision
Site Ecole Polytechnique de l’Université de Nantes
La Chantrerie — BP 50609 — 44306 Nantes cedex 3
{remi.lehn, henri.briand, fabrice.guillet}@univ-nantes.fr

Summary. A classical limit of association rule at the decider’s point of view is in

the combinatorial nature of the association rules, resulting in numerous rules. As the
overall quality of an association rule set can be considered as insight of the studied
domain given to the decider by the interpretation of the rules, too many rules can
make an harder interpretation then a worse quality of the overall process.
To get more readable rules and thus improve this global quality criterion, we
apply techniques initially designed for redundancy reduction in functional depen-
dencies sets to association rules. Although the two kinds of relations have different
properties, this method allow very concise representations that are easily understood
by the decider and can be further exploited for automatic reasoning.
In this paper, we present this method, compare it to other approaches and apply
it to synthetic datasets. We end with a discussion about the information loss resulted
of the simplification.

Key words: Minimal Covers, Closure, Interpretation of Association Rules, Deduc-

tive Reasonning

1 Introduction

The amount of collected data grows continuously. Decision tasks performed

must take this growth into account to deal with prediction, action evaluation
or validation, in the context of a large variety of application fields like man-
agement, profit optimization or analysis. The KDD (Knowledge Discovery in
Databases) area scopes this range of applications in the goal of providing auto-
mated tools and adapted data representations to help an expert user in finding
the evidences needed for the decision tasks. This assumes a human centered
KDD process. As a human centered process involving automated procedures,
it needs a targetted problem representations that are both realistic from the
user’s point of view and computable from a machine point of view.

R. Lehn et al.: User-System Interaction for Redundancy-Free Knowledge Discovery in Data,

Studies in Computational Intelligence (SCI) 127, 463–479 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
464 R. Lehn et al.

Among KDD techniques, association rules [2] allow the capture and the
representation of implicative patterns that tolerate a small set of counter-
examples —e.g. birds that cannot fly or sport cars that are not red. Associa-
tion rules can be enhanced with statistical evaluations and filters such as the
Intensity of Implication family of indices.
Association rule discovery is motivated by the exploitation of operational
databases to discover a new knowledge, that was unknown before the discovery
and that is potentially exploitable in a decision making process [19]. Many
performant algorithms have been published to optimize the association rules
search [8, 16] but they mainly focus on algorithmic optimization rather than
on knowledge usability.
One of the fundamental hypothesis of association rule discovery is that
the user does not specify the goal of the search. Because of the intrinsically
combinatorial nature of the search and the lack of the goals, the classical
use of these algorithms, chaining data selection, data formatting, frequent
sets induction, rules calculation and rule presentation to the user, generally
outputs quantities of rules, without order of any kind, which is in contradiction
with the principle of knowledge readability and usability for a decision process.
Experiments using a direct application of association rules algorithms like A
Priori, resulted in thousands of rules. We can then seriously contest the quality
of the vision of the studied domain provided by the association rules to the
user if he has to explore thousands of rules. We can contest as well the quality
of the induction itself if the energy that the user has to involve to interpret
the association rules is nearly the same as the energy he would have to deploy
to get the same domain understanding by directly browsing the database.
A classical answer to this problem is to set high thresholds on quality
indices that evaluate individual rules, to eliminate the least pertinent rules as
measured by these indices. But there are cases where this strategy cannot be
applied: when the user doesn’t know where to set thresholds corresponding
to the kind of knowledge he is looking for or when he’s looking for knowledge
with properties other than those matched by the available indices, or when
there are a lot of hidden dependencies in the data. Further global criteria have
been proposed, in addition to those measured on individual rules:
• operational criteria, in precise decision making tasks [6], but, these criteria
are specific and domain-dependent.
• Readability criteria, whose precise evaluation is based on a cognitive qual-
ification of the user’s perception of the represented knowledge. This cri-
terion depends of the visualization interfaces and their adaptation to the
decision making tasks. If we assume a linear, non-interactive, acquisition
of the knowledge by a decision maker, the limitation of the amount of rep-
resented rules, in association with a reading convention, is an important
factor of improvement of these criteria.
 User-System Interaction for Redundancy-Free KD in Data 465

• The exploitation of the rules for automated tasks such as inference engines.
In this case, specific properties of the rules for inference, for example the
respect of logical properties, are evaluated as knowledge quality.
To meet those criteria, we propose to limit the number of association
rules by not representing rules that can be inferred by the user himself in
a logical reasoning. So eliminated rules are then considered as redundant as
opposed to the other represented rules. This redundancy elimination considers
a global criterion that is a complement to the evaluation of the quality of each
individual rule.
The redundancy elimination strategy strongly depends on the theory of
the represented knowledge, including a definition of the inferences that can
be made by the user reading the knowledge representation. Several models
have been proposed and some of these models are coupled with discovery
algorithms [12,13,15,25]; other models are coupled with specific representation
models such as Galois Lattices [4,5,10,23,24,29], or with measures that allow
approximating frequent itemsets [1].
Our proposed representation model is based on logical properties with the
assumption of an implicative behavior of association rules. It is based on a
closed itemset algorithm (see Ceglar and Roddick, 2006 [8] for a definition of
this class of association rule mining algorithms), with a pure logical construc-
tion of the closure relation. This hypothesis is enforced with the assumption
that a user or an automated deduction system will make logical deductions
using the ruleset during the interpretation of the rules.
Efficient methods have been proposed for redundancy elimination in func-
tional dependencies sets and functional dependencies are known to support
logical properties [27]. Thus we apply one of these methods, the minimal cov-
ers, to association rules filtering. This filtering is very efficient as it gives very
minimal representations, but there exists some issues where association rules
do not respect the logical assumptions of our model and for which the redun-
dancy reduction gives over-generalized rules, that, while being in respect with
a logical reasoning, are in contradiction with some statistical measures of the
rules. We detail in this paper these cases, and show, using synthetic examples,
the information loss encountered using our method.

2 Redundant functional dependencies

Important work has been realized in the past for redundant functional depen-
dencies elimination in relational databases, for example Ullman in the early
80’s [3, 7, 9, 26, 27]; the results of this work have been directly embedded into
tools that allow the automatic production of a functional dependencies rep-
resentation from relational databases. Dep-Miner [20] is an example of such
a tool.
466 R. Lehn et al.

2.1 Definitions

Functional dependency and Armstrong’s axioms :

There exists a functional dependency (FD) A → B between a subset A of the

attributes of a relation R and another subset of attributes B if the relation R
associates one and only one set of values of attributes from B for each possible
combination of values of the attributes of A [21,27]. For example, the relation
R defined as:
R
a b c
a1 b1 c1
a2 b1 c2
a3 b2 c3
a3 b2 c4
holds a functional dependency a → b because each different value of a is
associated with one and only one value of b. This relation holds a functional
dependency c → ab too.
The three main axioms for calculus over functional dependencies systems
are the Armstrong’s axiom [28]:
Reflexivity : | > A ∪ B → A. (1)
Augmentation : A → B| > A ∪ C → B ∪ C. (2)
Transitivity : A → B, B → C| > A → C. (3)
A quite direct application of the augmentation axiom to association rules
was proposed by Toivonen et al., 1995 [25].

Closure on a FD set :

The closure on a FD set, F = (A, ∪, →), is defined on an attribute set A as

the set F + of FD that can be written from F by the repeated application of
Armstrong’s axioms (definition from Ullman, 1989 [28]). Two FD sets F1 and
F2 are equivalent if F1+ = F2+ . The closure of a subset X ⊂ A of attributes
over a FD set F = (A, ∪, →) is the set

XF+ = ∪i {Yi | (X → Yi ) ∈ F + } (4)

For example, the closure on the FD set F = {a → b, ab → c} is the set
F + = {a → a, ab → a, ac → a, abc → a, b → b, ab → b, bc → b, abc → b, c →
c, ac → c, bc → c, abc → c, ab → ab, abc → ab, ac → ac, abc → ac, bc →
bc, abc → abc, a → b, a → ab, ac → bc, ac → abc, ab → c, ab → ac, ab →
bc, a → c, a → ac, ab → abc}. Note that, among others, a → c is member of
the closure.
 User-System Interaction for Redundancy-Free KD in Data 467

2.2 Functional dependency decomposition

FD can be rewritten using union and decomposition as defined by the following

equivalence:

{A → B} ≡ {A → {b} | b ∈ B}. (5)

Functional dependencies decomposition consists in a FD set rewriting as
the right hand side of this equivalence (5). There are two advantages in this
rewriting: eliminating the redundancy dues to union and decomposition and
simplifying the processing of the FDs [7, 28]. At the decider’s point of view,
the writing of the FDs as the left hand side of the equivalence (5) gives a
more concise representation of the FDs and then has to be considered after
the redundant FD elimination.

3 Minimal covers
The minimal covers is the minimal FD set, F̂ computed from FDs F such
as F̂ + = F + and F̂ is minimal. F̂ is minimal if it does not contain neither
redundant FDs, nor superfluous attributes. A FD is said to be redundant if
it can be written using the Armstrong’s axioms on the FD system with the
exception of this FD: X → Y of F is redundant if F ⊂ (F \ {X → Y })+ . This
condition is satisfied if (X → Y ) ∈ (F \ {X → Y })+ . By using the definition
of the closure of an attribute set over a FD set (4), a FD can be qualified
+ +
as redundant if Y ⊂ X(F \{X→Y }) . Ullman [28] shows that if Y ⊂ XF , then
(X → Y ) ∈ F + .
An attribute x of the left hand side of a FD X → Y is superfluous if the
FD (X \ x) → Y can be computed using the Armstrong’s axioms on the FD
system, i.e.
+
F ⊂ ((F \ {X → Y }) ∪ {(X \ x) → Y }) . This condition is satisfied if ((X \
x) → Y ) ∈ F + or Y ⊂ (X \ x)+ F.
For example, the minimal covers of F = {a → b, ab → c, ac → d, a → c} is
F̂ = {a → b, ab → c, a → d}, because {a → b, ab → c} allows to infer a → c,
then c is superfluous in ac → d and a → c is redundant.

3.1 Proposed algorithms

The direct application of the previous definitions allows the computation of

the minimal covers. It is however non-deterministic at the state of the choice
of a FD to evaluate and the choice between eliminating a redundant FD or a
superfluous attribute [3]. The closure computation is a NP-hard problem, as
the rewritings of formulae only using the reflexivity axiom, producing n×2n−1
FDs given n attributes, solves the boolean satisfiability problem (SAT).
The closure computation can be avoided by providing a boolean function
that is true if an attribute subset is included in the closure of another attribute
468 R. Lehn et al.

subset over a FD set. Furthermore, the decomposition of the FDs giving FD

whose right-hand sides have only one attribute (5) allows to consider only
whether a single attribute belongs to the closure.

Determination of an attribute belonging to the closure of an

attribute set over a FD set :

The reflexivity axiom (1) can be rewritten into

| > (A ∪ A) → A
(6)
(A ∪ A) → A | > A → A;
therefore, y ∈ XF+ if y ∈ X; because x is in the right hand side of a FD
belonging to F + for which the left hand side is included into {x}.
Armstrong’s augmentation axiom (2) allows the rewriting of a FD (A →
B) ∈ F into

A→B | > (A ∪ A) → (A ∪ B)
(7)
and (A ∪ A) → (A ∪ B) | > A → (A ∪ B);
the addition of the FD ((A ∪ B) → C) ∈ F , the transitivity axiom (3)
allows to write

A → (A ∪ B) (7)
| > A → C. (8)
(A ∪ B) → C
The same rewritings can be achieved by the application of Armstrong’s
axioms on FDs whose left hand side is included in the attribute set A ∪ B.
The only demonstration of B ⊂ A+ F is enough to determine that if (A ∪ B →
C) ∈ F , then C ⊂ (A ∪ B)+F . Furthermore, there aren’t any other rewriting
starting from A → B giving FDs with right hand side that contain only
subsets of A ∪ B, then

y ∈ XF+ if ∃A → B ∈ F | A ⊂ X and y ∈ (X ∪ A)(F \{A→B}) . (9)

The determination of the belonging of an attribute to the closure of an

attribute set over a FDs set can be written as


y∈X
y∈ XF+ if (10)
or ∃A → B ∈ F | A ⊂ X and y ∈ (X ∪ A)+
(F \{A→B}) .

This recursive definition can be easily translated into an iterative form

(terminal recursion) giving the algorithm described by algorithm 1. The pro-
posed algorithms are based on a greedy algorithm (algorithmé2) : for each
step of this algorithm (lines 17 to 25) one or more rules are evaluated to up-
date the partial closure Xi (lines 22 to 23) or the end condition is reached.
 User-System Interaction for Redundancy-Free KD in Data 469

As rules are only evaluated once, the maximum number of iterations of the
closure main loop (lines 17 to 25) is the number of rules in the ruleset. Thus
this algorithm performs linearly with the number of rules of the computed
ruleset.

Data : A ruleset F .
Result : The minimal covers F̂ of F .
1 Let Fk = ∅;
2 foreach (X → Y ) ∈ F do
3 Let Xk = ∅;
4 foreach x ∈ X do
5 if Y 6⊂ (X \ x)+
F then
6 Xk = Xk ∪ {x};

7 if Xk 6= ∅ then
8 Fk = Fk ∪ {Xk → Y };

9 F̂ = Fk ;
10 foreach (X → Y ) ∈ Fk do
11 if Y ⊂ X(+F̂ \{X→Y }) then
12 F̂ = (F̂ \ {X → Y });

Algorithm 1: Minimal covers.

3.2 Examples

Figures 1, 2 and 3 show examples of the application of the proposed algo-

rithms. Line numbers in these examples refer to line numbers in algorithms 1
and 2.

4 Related work
4.1 Propositional logic

Armstrong’s axioms are theorems of the propositional logic [28]. It can then be
proven that every computable expression using Armstrong’s axioms are true
formulae for the propositional logic. Kaufman gave the proof that the theory
of FD redundancy is valid for logical implications [17], therefore, that a FD
system F = (A, ∪, →) shares its properties with a world in propositional logic
w = (A, ∧, →) where A is a set of propositions, ∧ is the logical conjunction
and → is the logical implication.
470 R. Lehn et al.

Data :• F : a ruleset.
• X : a set of attributes.
• y : an attribute.
Result : a boolean : true if y ∈ XF+ , false else.
13 Let Fi = F ;
14 Let Xi = X;
15 Let Fk = ∅;
16 closed = false;
17 while not closed and y 6∈ Xi do
18 closed = true;
19 Fk = ∅;
20 foreach A → B ∈ Fi do
21 if A ⊂ Xi then
22 Xi = Xi ∪ B;
23 closed = false;
else
24 Fk = Fk ∪ {A → B};
25 Fi = Fk ;
26 if y ∈ Xi then
27 y ∈ XF+ !;
else
28 y 6∈ XF+ !;

Algorithm 2: Determination if an attribute y belongs to the closure of an

attribute set X over a ruleset F .

This properties have been applied to association rules in several ap-

proaches [22, 25], using monotony properties of conjonctions of items for re-
dundancy elimination.

4.2 Conceptual lattices

Minimal covers and conceptual lattices share the use of inclusions between
extensions of the represented attribute combinations to limit the amount of
represented knowledge.

Association rules and Galois lattices :

Logical implications involved in the computation of the minimal covers are

captured by pseudo-intent (non-closed descriptions) are implicit and are not
represented on the Galois lattice but can be inferred by the user [23] (exam-
ple shown by figure 4). Galois lattices can represent the other, non-logical,
association rules. Furthermore, it has been proven that the representation of
a limited subset of association rules using a Galois lattice is sufficient for the
 User-System Interaction for Redundancy-Free KD in Data 471

1 superfluous attributes: F = a → b, a ∧ c → b
3 superfluous attributes: (X → Y ) = (a → b)
3 superfluous attributes: (X → Y ) = (a ∧ c → b)
5 superfluous attributes: (X \ x) = ({a, c} \ {a})
closure: b ∈ XF+ ?
closure: F = {a → b, a ∧ c → b}, X = c
15 closure: Fk = ∅
20 closure: (A → B) = (a → b)
24 closure: A 6⊂ Xi
closure: Fk = {a → b}
closure: Xi = {c}
20 closure: (A → B) = (a ∧ c → b)
24 closure: A 6⊂ Xi
closure: Fk = {a → b, a ∧ c → b}
closure: Xi = {c}
28 closure: b 6∈ XF+ !
5 superfluous attributes: a 6∈ ({c})+ F
5 superfluous attributes: (X \ x) = ({a, c} \ {c})
closure: b ∈ XF+ ?
closure: F = {a → b, a ∧ c → b}, X = a
15 closure: Fk = ∅
20 closure: (A → B) = (a → b)
22 closure: A ⊂ Xi
closure: Fk = {}
closure: Xi = {a, b}
20 closure: (A → B) = (a ∧ c → b)
24 closure: A 6⊂ Xi
closure: Fk = {a ∧ c → b}
closure: Xi = {a, b}
25 closure: Fi = Fk
closure: Fi = {a ∧ c → b}
27 closure: b ∈ Xi !
closure: b ∈ XF+ !
5 superfluous attributes: c ∈ ({a})+ F
5 superfluous attributes: c is superfluous.
8 superfluous attributes: (Xk → Y ) = (a → b)

Fig. 1. Example of superfluous attributes elimination over the set of rules {a →

b, a ∧ c → b}.

inference of the whole association rule system [10, 23, 29] with the following
reading conventions:
1. the support of a non-closed description (pseudo-intent) which is not rep-
resented on the Galois lattice is equal to the support of the closed set in
which it is included; in the example given figure 4, the support of a (non
closed) is the same as the support of a ∧ b ∧ c (closed). [10, 23];
472 R. Lehn et al.

9 redundant FD: F̂ = Fk = a → b, a → c, b → c
11 redundant FD: (X → Y ) = (a → b)

11 redundant FD: F̂ \ (X → Y ) = a → c, b → c
closure: b ∈ XF+ ?
closure: F = {b → c, a → c}, X = a
15 closure: Fk = ∅
20 closure: (A → B) = (b → c)
24 closure: A 6⊂ Xi
closure: Fk = {b → c}
closure: Xi = {a}
20 closure: (A → B) = (a → c)
22 closure: A ⊂ Xi
closure: Fk = {b → c}
closure: Xi = {a, c}
25 closure: Fi = Fk
closure: Fi = {b → c}
15 closure: Fk = ∅
20 closure: (A → B) = (b → c)
24 closure: A 6⊂ Xi
closure: Fk = {b → c}
closure: Xi = {a, c}
28 closure: b 6∈ XF+ !
11 redundant FD: b 6∈ ({a})+ (F̂ \(X→Y )
!

12 redundant FD: F̂ = a → b, a → c, b → c
11 redundant FD: (X → Y ) = (b → c)

11 redundant FD: F̂ \ (X → Y ) = a → b, a → c
closure: c ∈ XF+ ?
...
(in a similar way, it can be proven that:)
11 redundant FD: c 6∈ ({b})+ (F̂ \(X→Y )
!

12 redundant FD: F̂ = a → b, a → c, b → c

Fig. 2. Example of redundant rule elimination over the ruleset {a → b, b → c, a →

c} : examination of the rules a → b and a → c.

2. every description on the frontier of the search1 is closed [23];

3. every association rules and theirs confidences can be inferred using the
represented intents of the inferred pseudo-intents [23].
Galois lattices however need the computation and the representation of the
whole lattice -a non represented description means a non closed description-
where the representation of the minimal covers allows correct inferences. In
the previous example (figure 4), if we only represent b → b ∧ c and c → b ∧ c
it is impossible to infer a → b ∧ c as the information about the closure of a is
not represented. Another limit to Galois lattices is when there are no logical

1
The frontier of the search is the set of the most specific intents.
 User-System Interaction for Redundancy-Free KD in Data 473

11 redundant FD: (X → Y ) = (a → c)

11 redundant FD: F̂ \ (X → Y ) = a → b, b → c
closure: c ∈ XF+ ?
closure: F = {a → b, b → c}, X = a
15 closure: Fk = ∅
20 closure: (A → B) = (a → b)
22 closure: A ⊂ Xi
closure: Fk = {}
closure: Xi = {a, b}
20 closure: (A → B) = (b → c)
22 closure: A ⊂ Xi
closure: Fk = {}
closure: Xi = {a, b, c}
25 closure: Fi = Fk
closure: Fi = {}
27 closure: c ∈ Xi !
closure: c ∈ XF+ !
11 redundant FD: c ∈ ({a})+ F̂ \(X→Y )
!
redundant FD: a → c is redundant !

12 redundant FD: F̂ = a → b, b → c

Fig. 3. Example of redundant rule elimination over the ruleset {a → b, b → c, a →

c} : examination of the rule b → c.

implication. In this case, every discovered description is closed, then there are
no pseudo-intent and then the Galois lattice is equivalent with the description
inclusion lattice.
Algorithms are proposed to compute frequent closed itemsets [29], as a
replacement for the original frequent itemsets step in the classical association
rule discovery process.

δ-free sets :

Boulicaut et al. [5] proposed a new notion, the δ-free descriptions, that meet
this latter problem. It is an enhancement of the closure definition, taking into
account quasi-inclusions2 . A description is said to be δ-free if there aren’t
any rule between subsets of these description, that is invalidated by at most
δ examples. The δ factor is supposed to be small. The set of δ-free frequent
descriptions allows to approximate the set of the descriptions, with an error
rate defined by support and confidence thresholds. It reduces the number of
represented descriptions while reducing the time required for their computa-
tion.

2
that can be seen as extents statistical implications.
474 R. Lehn et al.

1 Relation : 4 Galois lattice :

a b c b
X X X o b^c a^b^c
X X c

X 3 minimal covers : a−>b^c

X

2 33 logical implications :
a b c a^b a^c b^c a^b^c left hand sides
a X X X X
b X X X X X X
c X X X X X X
a^b X X X X
a^c X X X X
b^c X X X X X
a^b^c X X X X

right hand sides

3 The minimal covers of the initial 33 rules 2 can be used to infer the logical
rules 2 .
4 The conceptual lattice of the relation can be used to infer the whole set of
rules, including the minimal covers. In this example, to infer the minimal
covers, we use the following reasoning: an attribute a appears in the closed
set a ∧ b ∧ c, but it does not appear on any other represented closed set.
It means that descriptions a, a ∧ b and a ∧ c, which are described by a are
described by b and c too and then that a → b ∧ c.

Fig. 4. Comparison between a minimal covers and a Galois lattice.

5 Experimental results on the extension of the minimal

covers to association rules

Armstrong axioms are not valid for every kind of implicative systems where
statistical implications are considered. The figure 5 is an illustration of the
limits of the transitivity axiom for statistical implications.
There are the same limits for the augmentation axiom.

5.1 Further propositions to circumvent these limits

The use of logical propositions for association rules are interesting in three
categories of applications:
 User-System Interaction for Redundancy-Free KD in Data 475
C

A
B

Note : ( B , . . . ) here represents the set of objects for which a is true

A

(b, . . . , respectively)
1 Both statistical implications a → b and b → c are observable, as well is
observed the statistical implication a → c.
⇒ valid transitivity !

A
B

2 a → b and b → c are observable, but no relation exists between a and c

⇒ no transitivity !

A
B

3 a → b and b → c are observable, but the statistical implication a → ¬c is

observable
⇒ anti-transitivity !

Fig. 5. Limits of the axiom of transitivity for statistical implications.

476 R. Lehn et al.

1. the hypothesis that association rules behave like logical propositions can
be investigated with the validation of an expert. This is required to use
the rules as a knowledge base for inference engines of an expert system.
2. Setting thresholds on quality indices of the rules (e.g. a low support and
an high confidence) aims at getting a behavior of association rules that
is near logical propositions. There isn’t unfortunately any formula giving
the correct thresholds that will ensure a complete logical behavior, except
the trivial case of a confidence threshold of 1.
3. The redundancy elimination can be considered as a part of an interactive
process, providing the user a way to infirm or confirm his hypothesis dur-
ing his reasoning [18]. This interaction can be useful in finding exceptions
to the logical behavior of rules or sets of rules.
The first two points are confirmed by experiments whose results are pre-
sented by figure 6 and table 73 . These represents experiments on 100 rulesets,
varying min_conf 4 in {0.8, 0.9, 1} and ϕ5 in {0.8, 0.9, 1}. For each ruleset, we
check the minimal covers and the closure with the same quality criteria to de-
termine the ratio of valid inferences. Here, we assume that the user is able to
infer the closure himself from either the original ruleset or its minimal covers.
These experiments show that the higher the min_conf threshold is set, the
better the minimal covers and inferences on the basis on the minimal covers
are correct. A similar but weaker behavior is observed with the intensity of
implication.

6 Conclusion

We applied to association rules a redundancy elimination method for func-

tional dependency systems. This method eliminates redundant rules, that can
be inferred by a user in a logical reasoning. Properties of functional dependen-
cies are an example of rewriting rules that are both useful for propositional
logic and useful for human interpretation for data model design. We showed
that this method can strongly reduce the quantity of rules exposed to the user,
at the price of some approximation: there is no practical way for the user to
compute the quality indices associated with each individual inferred rule, and
the user can be disoriented to infer wrong rules, that does not exist in the
initial rule set, according to the quality criteria set during the rule discovery.
However the conciseness of the achieved representation, the resort on sim-
ple and usual logical properties, and the existence of efficient, polynomial,
algorithms for the minimal covers computation make this method useful for a

3
The program used to present these results can be downloaded at the following
URL : http://www.fc.univ-nantes.fr/~remi/felix/min-covers.
4
threshold on confidence.
5
Here, the original definition of intensity of implication is used [14].
 User-System Interaction for Redundancy-Free KD in Data 477

100
% rules in minimal covers / total discovered
% valid rules in minimal covers
90 % valid rules among inferred rules

80

70

60
% rules

50

40

30

20

10

0
0 500 1000 1500 2000 2500 3000 3500
number of rules

This vertical rule represents one experiment on the basis

of 3072 rules.
The minimal covers has 28 rules (0.91%) (+)
50 % of these 28 rules are valid ones (x)
73 % of the rules of the closure are valid (*)

Fig. 6. Sample experimental results

first acquisition of rules by a user, before starting an interactive mining among

the rules.
This method is also an alternative for the quality evaluation of a set of
rules, considering the global quality of a rule set used for reasoning and deci-
sion making, in addition to the sum of the qualities of individual rules.

References
1. F.-N. Afrati, A. Gionis, and H. Mannila. Approximating a collection of frequent
sets. In Proceedings of the Tenth ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, pages 12–19. ACM, 2004.
2. R. Agrawal, H. Mannila, R. Srikant, H. Toivonen, and A. Inkeri Verkamo. Fast
discovery of association rules. In Fayyad et al. [11], pages 307–328.
3. J. Atkins. A note on minimal covers. SIGMOD RECORD, 17(4):16–21,
December 1988.
478 R. Lehn et al.

min _conf percentage of successful experiments, in term of confidence

of rules inferred from the minimal covers.
0.8 66.1%
0.9 70.5%
1 98.1%
min _ϕ ∈ {0.9, 0.95, 1}

min _ϕ percentage of successful experiments, in term of intensity of

implication of rules inferred from the minimal covers.
0.9 72.4%
0.95 73.3%
1 76%
min _conf ∈ {0.8, 0.9, 1}
Note : here ϕ represents R. Gras’ intensity of implication [21].

Fig. 7. Experimental results : confirmation of the inferred rules using tests with
quality measures.

4. Y. Bastide, N. Pasquier, R. Taouil, G. Stumme, and L. Lakhal. Mining minimal

non-redundant association rules using frequent closed itemsets. In Proceedings
of the First International Conference on Computational Logic - CL2000, LNCS
1861, pages 972–987. Springer, 2000.
5. J.-F. Boulicaut, A. Bykowski, and C. Rigotti. Approximation of frequency
queries by means of free-sets. In Proceedings of Principles of Data Mining
and Knowledge Discovery, Lecture Notes in Computer Science, pages 75–85.
Springer-Verlag, 2000.
6. J.R. Brachman and T. Anand. The process of knowledge discovery in databases:
a human-centered approach. In Fayyad et al. [11], pages 37–58.
7. H. Briand, J.B. Crampes, Y. Hebrail, D. Herin Aime, J. Kouloumdjian, and
R. Sabatier. Les systèmes d’information. éditions DUNOD, 1986.
8. Aaron Ceglar and John F. Roddick. Association mining. ACM Comput. Surv.,
38(2):5, 2006.
9. C. Delobel and M. Adiba. Bases de données et systèmes relationnels. DUNOD
Informatique, 1982.
10. L. Dumitriu, C. Tudorie, E. Pecheanu, and A. Istrate. A new algorithm for
finding association rules. In Proceedings of Data Mining, volume 2, pages 195–
202. Wessex Institute of Technology, WIT Press, 2000.
11. U.M. Fayyad, G. Piatetsky-Sapiro, and P. Smyth, editors. Advances in Knowl-
edge Discovery and Data Mining. AAAI Press, 1996.
12. L. Fleury. Adaptation d’une méthode de recherche de la couverture minimale
d’un ensemble de dépendances fonctionnelles pour l’élimination des redondances
dans un système de règles. INFORSID, Aix en Provence, 1994.
13. J.-G. Ganascia. AGAPE et CHARADE : deux techniques d’apprentissage sym-
bolique appliquées à la construction de bases de connaissances. PhD thesis,
Université de Paris Sud, 1987.
14. R. Gras, S. Almouloud, M. Bailleuil, A. Lahrer, M. Polo, H. Ratsimba-Rajohn,
and A. Totohasina. L’implication statistique : nouvelle méthode exploratoire de
données. Edition de la pensée sauvage, 1996.
 User-System Interaction for Redundancy-Free KD in Data 479

15. J.-L. Guigues and V. Duquennes. Familles minimales d’implications informa-

tives résultant d’un tableau de données binaires. In Mathématiques et sciences
humaines, number 95, pages 5–18. 1986.
16. J. Hipp, U. Güntzer, and G. Nakhaeizadeh. Mining association rules: deriving
a superior algorithm by analysing today’s approaches. In Proceedings of the 4th
European Conference on Principles of Data Mining and Knowledge Discovery,
volume 1910 of Lecture Notes in Computer Science, pages 159–168. Springer
Verlag, 2000.
17. A. Kaufmann. Nouvelle logique pour l’intelligence artificielle. Mathématiques
appliquées. Editions Hermes, 1987.
18. P. Kuntz, R. Lehn, and H. Briand. A user-driven process for mining association
rules. In Proceedings of Principles of Data Mining and Knowledge Discovery,
volume 1910 of Lecture Notes in Computer Science, pages 483–489. Springer-
Verlag, 2000.
19. R. Lehn, F. Guillet, P. Kuntz, H. Briand, and J. Philippé. Felix: An interactive
rule mining interface in a kdd process. In Proceedings of the 10th Mini-Euro Con-
ference, Human Centered Processes, HCP’99, pages 169–174. Ecole Nationale
Supérieure des Télécommunications de Bretagne, 1999.
20. S. Lopes, J.-M. Petit, and L. Lakhal. Efficient discovery of functional dependan-
cies and armstrong relations. Rapport de recherche LIMOS, Université Blaise
Pascal, Clermont-Ferrand II, 1999.
21. H. Mannila and K.-J. Räihä. The Design of Relational Databases. Addison-
Wesley, 1992.
22. B. Padmanabhan and A. Tuzhilin. On characterization and discovery of minimal
unexpected patterns in rule discovery. In IEEE Transactions on Knowledge and
Data Engineering, volume 2, pages 202–216. IEEE Press, 2006.
23. N. Pasquier. Data Mining : Algorithmes d’Extraction et de Réduction des Règles
d’Association dans les Bases de Données. PhD thesis, Université de Clermont-
Ferrand II, 2000.
24. N. Pasquier, Y. Bastide, R. Taouil, and L. Lakhal. Efficient mining of association
rules using closed itemset latices. Information Systems, 24(1):25–46, 1999.
25. H. Toivonen, M. Klemettinen, P. Ronkainen, K. Hatonen, and H. Mannila. Prun-
ing and grouping of discovered association rules, 1995.
26. J.D. Ullman. Principles of Database Systems. Computer Science Press, 1982.
27. J.D. Ullman. Principles of Database and Knowledge–base Systems, volume 1.
Computer Science Press, 1989.
28. J.D. Ullman. Reasoning about functional dependencies, chapter 7.3, pages 382–
392. Volume 1 of [27], 1989.
29. Mohammed Javeed Zaki. Generating non-redundant association rules. In
Knowledge Discovery and Data Mining (KDD’00), pages 34–43. ACM Press,
2000.
Fuzzy Knowledge Discovery Based on Statistical
Implication Indexes

Maurice Bernadet

LINA / Ecole Polytechnique de l’Université de Nantes,

Rue Christian Pauc, La Chantrerie,
BP 60601, 44306 Nantes Cedex 3, France
[email protected]

Summary. We describe one application of statistical implication indexes to fuzzy

knowledge discovery. After recalling principles of fuzzy logics, we explain how we
have adapted statistical indexes to fuzzy knowledge: the support, the confidence
and a less common index, the intensity of implication. These indexes highlight sta-
tistical links between conjunctions of fuzzy attributes and fuzzy conclusions, but do
not evaluate the associated fuzzy rules, which depend of the chosen fuzzy operators.
Since fuzzy operators are numerous, we evaluate their sets by applying the gener-
alized modus ponens on the database and by comparing its results to the effective
conclusions. We give a summary of the results on several databases, and we present
the sets of fuzzy operators that appear to be the best. Studying methods to aggre-
gate fuzzy rules, we show that in order to keep classical reduction schemes, fuzzy
operators must be chosen differently. However, one of these possible operator sets is
also one of the best for processing the generalized modus ponens.

Key words: Statistical implication, fuzzy knowledge discovery, fuzzy implication,

fuzzy operators.

1 Introduction

Fuzzy logics are extensions of classical logics, which allow intermediate truth-
values between True and False [31]. They may express knowledge in a more
natural way than classical Boolean logics, allowing graduated attributes as
in the sentence “X is rather high” (for “X is high” is rather true) and then
assigning, for instance, a truth value of 0.8 to the proposition X is high. Fuzzy
logics offer many logical operators, which permits a good translation of various
kinds of knowledge. In the domain of knowledge discovery [13], considering
that crisp intervals on continuous attributes are difficult to interpret and that
strict thresholds are often too abrupt, one may think that fuzzy logics could
improve the expressiveness of extracted knowledge.

M. Bernardet: Fuzzy Knowledge Discovery Based on Statistical Implication Indexes, Studies

in Computational Intelligence (SCI) 127, 481–506 (2008)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
482 M. Bernardet

Some methods, specific to fuzzy knowledge discovery, have already been

defined, allowing to extract fuzzy rules; these methods are often holistic: they
may use neural networks [10, 17, 18] or genetic algorithms [8, 23, 26]. However,
we have preferred to adapt classical knowledge discovery methods to fuzzy
knowledge because classical methods are more analytical, hence allowing a
follow-up of the mechanisms during their progress. The methods we have
defined extract interesting fuzzy rules by computing statistical indexes to
evaluate statistical implications between fuzzy premises and fuzzy conclusions.
Once statistical implications have been highlighted, they are used to build
fuzzy rules based on fuzzy operators, which must be chosen according to the
considered application.
We here summarize a significant part of our work on this topic, examining
it in the perspective of its relation to statistical implication analysis. We have
added recent results about fuzzy rule reduction, which shows one important
difference between fuzzy rules and statistical rules: while generally statistical
rules may not be aggregated, a proper choice of fuzzy operators allows the use
of classical rule reduction schemes.
The first part describes the initial process of our fuzzy knowledge discovery
methods, which requires a definition of fuzzy partitions to convert numerical
or symbolic attributes into fuzzy ones. The second part justifies the classical
indexes we have chosen to evaluate the statistical implications; these indexes
are the classic support and confidence of a rule, associated with a less common
index, the intensity of implication. We then explain how we have adapted these
indexes to fuzzy attributes and we describe one algorithm we use to explore
the set of possible rules.
Such exploration algorithms, based on statistical indexes, highlight statis-
tical links between conjunctions of fuzzy attributes and fuzzy conclusions, but
they do not evaluate the associated fuzzy rules, which depend on the chosen
fuzzy operators. Since fuzzy logics offer numerous possibilities to define logical
operators, the next paragraph describes the methods we use to evaluate sets of
fuzzy operators. These methods are based on another statistical index, which
evaluate fuzzy rules, given one set of fuzzy operators; we call this index the
GMP-pertinence (GMP stands for generalized modus ponens). The next two
parts describe the results given by this method on a simplified example and
summarize the results on real databases found in the UCI repository. We then
review the best sets of fuzzy operators found with this method. Subsequently,
we have studied sets of fuzzy operators which allow easy reduction of fuzzy
rules. We have highlighted that most of these sets are not the most GMP-
pertinent; one must distinguish two kinds of applications for which the chosen
fuzzy operator sets should be different: on the one hand, there are knowledge-
based systems using the GMP for decision aids, and on the other hand, there
are analytical systems to help experts to explore knowledge extracted from
databases. However, associating Gödel-Brouwer’s implication with standard
Zadeh’s minimum and maximum appears as a good compromise.
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 483

2 The “Fuzzification” Process

Translation of classical attributes, numeric or symbolic, into fuzzy ones is
called “fuzzification”. To carry out this translation we need at first to define
pseudo fuzzy partitions, allowing a classical attribute to belong to several fuzzy
classes (generally at most two), with degrees less than 1 and which usually
sum up to 1. Once these pseudo fuzzy partitions have been defined, one may
translate the values of each classical attribute into its degrees of membership
to the different modalities of the corresponding fuzzy attribute.

2.1 Definition of Fuzzy Partitions

Let us first recall that fuzzy logics evaluate the truth-value of a fuzzy propo-
sition “X is A”, as the degree to which X belongs to the fuzzy set A: if µA (X)
is the membership (or characteristic) function of the fuzzy set A, one may
write Truth(“X is A”) = µA (X),
Fuzzy sets allow the definition of fuzzy C-partitions or “pseudo partitions”
in which each value of a continuous attribute may be classified into several
fuzzy classes, with a total membership of 1. These fuzzy pseudo partitions
allow the conversion of continuous attributes into fuzzy ones, then giving
the truth-value of fuzzy propositions. For a continuous attribute CA, varying
from minCA to maxCA, one can define a fuzzy pseudo partition in several
ways [7, 22].
The simplest method divides the interval [minCA, maxCA] in n sub-
intervals, with a small percentage of coverage between two adjacent ones,
giving each sub-interval a symbolic name related to their position. For in-
stance, one may divide the interval [minCA, maxCA] in 5 sub-intervals with
an overlap of 20%, then giving 5 fuzzy modalities for this attribute, such as
strong negative, rather negative, medium, rather positive and strong positive
(Fig. 1).

Fig. 1. A fuzzy C-partition (α =1, strong negative; α =2, rather negative;

α = 3, medium; α = 4, rather positive; α =5, strong positive).
484 M. Bernardet

The fuzzy classes may also be defined by experts; otherwise one may choose
3 or 5 classes as standard options. Different numbers of classes may also
be used, but too high a number of classes might too heavily slow down the
knowledge discovery process. It is often interesting to try several numbers of
classes because it is difficult a priori choosing the number of classes that will
give a good partition is difficult.
Another kind of method extracts the number of classes and defines the
fuzzy C-classes from the database. These methods consider values of the at-
tributes giving the same conclusion and, whenever possible, cluster these val-
ues into the same fuzzy sets, with a membership value equal to the rate of
samples giving this conclusion. These methods often use histograms of at-
tribute values for each possible conclusion. Moreover, it is possible to develop
a more satisfactory method, by generalizing optimal discretization methods
such as those studied in [33] to fuzzy logics; we have recently defined such
a method, based on clustering, which gives more satisfying results than the
previous ones, but with the drawback of needing more computing time [5].

2.2 Fuzzification of a Database

Once the fuzzy classes have been defined for each attribute, one may convert
the related values of each item by mapping these values to the membership val-
ues of each fuzzy class associated to the considered classical attribute (Fig. 2).

Fig. 2. Mapping from the value V of the continuous attribute CA into membership
values of fuzzy attributes (here, only µ3 and µ4 are non zero).

3 Choosing Statistical Implication Indexes

Several indexes may be used to evaluate classical rules [1], and we have chosen
three of them [3, 6]: the confidence, the support and a less known index, the
intensity of implication. Let us consider two propositions a and b associated
respectively with A and B, the sets of elements which verify them.
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 485

• The confidence of a rule such as “if a then b” expresses the conditional

probability of b to be true when a is true; calling na∧b , the number of
items verifying “a and b” and na the number of items verifying a, the
confidence may be evaluated by
na∧b
Conf idence(a ⇒ b) = (1)
na

• The support of a rule “if a then b” may be defined as the rate of occurrences
of items verifying “a and b”, related to all items of the database; calling
na∧b the number of items verifying “a and b” and nE the total number of
items in the database, the support may be evaluated by
na∧b
Support(a ⇒ b) = (2)
nE

• The intensity of implication is an index expressing the quality of a rule.

This index, defined by R. Gras and A. Larher [15], is based on simple
probability concepts: since the cardinalities of two subsets A and B of a
reference set E are determined by the objects of the database belonging to
A and B, we consider two random subsets X and Y having respectively the
same cardinalities as A and B. The implication a ⇒ b is characterized by
the relation A ⊂ B and its counter-examples are associated to the subset
A ∩ B. So, we compare the cardinality of A ∩ B (given by the database)
to the random variable given by the cardinality of A ∩ B, supposing that
there is no statistical link between X and Y (Fig. 3).

Fig. 3. X and Y vary randomly in E

If the cardinality of A ∩ B is unusually small compared to the expected

value of the distribution of the cardinalities of X ∩ Y , we consider a ⇒ b as a
quasi-implication. The intensity of a ⇒ b is therefore the difference between
486 M. Bernardet

1 and the probability for the random variable “cardinality of X ∩ Y ” to be

smaller than the cardinality of A ∩ B. Intuitively, the intensity of implication
measures the degree of statistical astonishment at the size of A∩B, considering
the sizes of A, B and E, and assuming there is no a priori link between A and
B.
One may note that the confidence does not answer the question: “What
probability is there to have an implicative link between propositions a and b?”.
A conditional probability rather answers the question “What is the probability
of proposition b when proposition a is true? ” So, these two measures are
complementary: in a learning approach, the intensity of implication allows
the withdrawal of little pertinent implications, while conditional probabilities
give for each rule an inference mechanism for uncertain reasoning in an expert
system. Therefore, the quality of one implication is better when the number
of its counter-examples is smaller than their expected number, that is to say
when the quantity P (Card(X ∩ Y ) ≤ Card(A ∩ B)) is small.
Consequently, it is the observation of the “smallness of Card(A ∩ B) com-
pared to Card(X ∩ Y )” which is taken as a basic evaluator of the interest of
the quasi-implication a ⇒ b. The intensity of implication is then defined by
the function


ϕ(a, b) = 1 − P Card(X ∩ Y ) ≤ Card(A ∩ B) (3)

Let us call n = Card(E), nA = Card(A), nA = Card(A), nB =

Card(B), nB = Card(B), nA∩B = Card(A ∩ B), nA∩B = Card(A ∩ B);
the random variable Card(X ∩ Y ) obeys a hypergeometric distribution:
n−n −k n −k

CnkA · Cn−nAB CnkA · CnAB
P Card(X ∩ Y ) = k = = n (4)
Cnn−nB Cn B
and
Card(A∩B) n −i

X Cni A · CnAB
P Card(X ∩ Y ) ≤ Card(A ∩ B) = n (5)
Cn B
i=0
i ≥ nA − nB

A detailed comparison of the intensity of implication with other statistical

indexes has been made in [9] and [12]. Let us summarize these studies:
– the value of the intensity of implication increases with the size of the
learning set, while other indexes remain constant,
– the intensity of implication reflects the human way of withdrawing a pre-
vious opinion: some new counter-examples for a strong implication do not
change much this index, but progressively doubts appear, and finally a few
more counter-examples cause the withdrawal of this opinion;
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 487

– for similar reasons, the use of intensity of implication is well adapted to

noisy data since a small number of counter-examples do not necessarily
invalidate the implication;
– finally, the intensity of implication prohibits the generation of rules such
as a ⇒ b when the proposition b is true for nearly all examples of the
learning set: in that case it is not surprising that the set of examples for
which a is true is nearly included in the set of examples for which b is true.

4 Generalization of Statistical Indexes to Fuzzy

Knowledge

We then have generalized these three indexes to fuzzy knowledge, to search

statistical implicative links between fuzzy attributes. To allow fuzzy infer-
ences, each discovered link must be associated to the corresponding fuzzy
rule. However, results from [27] and [14], show that statistical implications
give more satisfying semantical results than two main fuzzy implications. That
is due to the fact that fuzzy implications generalize classical implications such
as p ⇒ q; such implications are true when p is false, whatever the value q. In
the same situation, statistical analysis reasonably judges that there is no link
between p and q. Nevertheless, since we not only wanted to study the seman-
tical aspects of implications, but also wanted to study inference mechanisms
on extracted rules, we have studied fuzzy implications to use with the main
inference mechanism of fuzzy logics, the Generalized Modus Ponens.
To handle fuzzy knowledge, machine learning indexes may exclusively be
based on the theory of fuzzy sets, like those described in [22] or in [2]; but
these indexes may as well be a generalization of indexes developed in classical
logic, like those described in [25, 28] or [32]. We have adopted this second
method and we have generalized the three classical indexes we have retained
to fuzzy knowledge; for this purpose we have applied a definition given by
Zadeh [30] for the probability of a fuzzy event.
Considering a set of objects E, with n domains of reference D1 , D2 , . . . , Dn ,
on which we define fuzzy attributes for objects in E, the set of classical propo-
sitions D becomes the set of all fuzzy propositions that can be expressed on
objects of E. The number of elements of E that satisfy a proposition p associ-
ated to the fuzzy set Pe with the characteristic function µP , may
P be evaluated
by the crisp cardinality of the fuzzy set Pe on E: Card(Pe) = µP (x).
x∈E
This notion of cardinality has recently been criticized: when a strong pro-
portion of items has a low membership in a fuzzy set, comparisons using this
cardinality can lead to absurd results [11]. This problem should not occur
within our approach, because the fuzzy subsets we use constitute a pseudo-
partition, built by consultation of an expert or by a clustering method, and
because these modes of construction produce fuzzy sets with a kernel (the set
488 M. Bernardet

of elements with a membership of 1) covering a large part of the support (the

set of elements with a membership greater than 0). However, to prevent a bad
partitioning, we verify that the support of each fuzzy set is not larger than
a threshold σ of the percentage of its kernel. If this is not true, we take into
account only the degrees of membership superior or equal to 0.5 (also called
the alpha cut to level 0.5) and we use as the cardinality of a fuzzy set Pe on a
P
referential E, a formula inspired of [24] and [29]: Card(Pe) = µP (x) .
x∈E, µP (x)≥0.5

More often, however, we use Zadeh’s crisp cardinalities of fuzzy sets. The
confidence of a rule, its support and its intensity of implication are then ex-
pressed by the same formulas as above, by replacing cardinalities of crisp sets
by cardinalities of fuzzy sets. Thus, if one calls > (t-norm) the fuzzy “and”
operator with a fuzzy complement µA(x) = 1 − µA(x), one can write:
X
nA = Card(A) = µA(x), (6)
x∈E
X X
nA = Card(A) = µA(x) = (1 − µA(x)), (7)
x∈E x∈E
X X
nB = Card(B) = µB(x) = (1 − µB(x)), (8)
x∈E x∈E
X X
nA ∩ B = Card(A ∩ B) = µA ∩ B(x) = >(µA(x), µB(x))), (9)
x∈E x∈E
X X
nA∩B = Card(A∩B) = µA ∩ B(x) = >(µA(x), (1 − µB(x))) (10)
x∈E x∈E

The confidence of the rule may still be expressed by

nA∩B
Confidence(X is A ⇒ Y is B) = , (11)
nA
Its support by
nA∩B
Support(X is A ⇒ Y is B) = , (12)
nE
And its intensity of implication by


ϕ(a, b) = 1 − P Card(X ∩ Y ) ≤ Card(A ∩ B) (13)
with
Card(A∩B) n −i

X Cni a · Cnab
P Card(X ∩ Y ) ≤ Card(A ∩ B) = n (14)
Cn b
i=0
i ≥ na − nb
Once the indexes to evaluate rules have been chosen, a fuzzy knowledge extrac-
tion algorithm may be used, with the same principles as a classical knowledge
extraction algorithm. The algorithm we use is an exploratory search in the
tree of possible rules.
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 489

5 A Knowledge Extraction Algorithm

This algorithm builds all the rules that can be constructed from a set of propo-
sitions, computes their confidence, their support and their intensity of impli-
cation and keeps the rules for which these indexes are above three respective
thresholds. To limit the number of rules studied and the exploration depth,
we also restrict to a maximum the number of propositions in the premises of
a rule. So, we use 4 thresholds α, β, γ, δ and one rule is kept if its confidence
is greater than α, its support greater than β, its intensity of implication over
γ and if its premises have at most δ propositions. The thresholds α, β, γ, δ
are chosen by the users in accordance to the number of rules that they wish
to obtain.
Rules are structured in a tree; the root of this tree (level 0) is the “rule
with the empty premises”, level 1 has rules using only 1 proposition in their
premises, . . . , level i has rules using i propositions in their premises, and so
on. The algorithm uses a depth first strategy; this search is not carried deeper
when the current rule has not the minimal support γ, or when the size of the
current rule is above δ.
Let us call
- c, the confidence of a rule, which must be greater than the threshold α;
- s, the support of a rule, which must be over the threshold β;
- i, the intensity of implication, which must be greater than the threshold
γ;
- l, the length of the rule (the number of propositions in its premises, which
must not be more than the threshold δ);
- E = e1 , e2 , . . . , en , the learning set;
- P = p1 , p2 , . . . , pn , the set of propositions describing examples in E;
- C, the set of propositions associated to the conclusions;
- D = a1 , a2 , . . . , am , the set of attributes in the possible propositions of
the premises;
- Fdecision , the fuzzy partition associated to the attribute of the classifying
decision;
- nFdecision , the cardinality of this partition;
- R, the set of rules produced.
Our algorithm, described below, uses two scanning procedures:
-“Forward” adds, when possible, a fuzzy proposition not used yet at this level
to the premises of the rule,
-“Backward” removes the last fuzzy proposition of the premises (the one on
the further right), and, if possible, replaces it by the following one. When this
proposition has no following one (because it uses the last modality of the last
attribute), the new most to the right proposition is removed and replaced, if
possible, and so on. When there are no more propositions in the premises, the
tree of premises has been completely explored and the algorithm ends.
490 M. Bernardet

Algorithm: Knowledge Extraction

Begin
R = ∅;
For all values vi ∈ Fdecision do
Let T , the tree of rules with the conclusion {adecision = vi };
Let B, the set of observations in E with {adecision = vi } true;
CurrentNode = Forward(T , R);
While the tree T has not been totally searched do
Let r : P remise ⇒ pi ∈ C, the rule associated to CurrentNode;
Let A, the set of examples in E in which Premise is true;
Compute c, s, i from the cardinalities of E, A and B;
Let l, the length of Premise ;
If(c ≥ α) and (s ≥ β) and (i ≥ γ) and (l ≤ δ)
Then R = R ∪ { P remisse ⇒ pi };
End If
If (s < β) or (l > δ) or (CurrentNode terminal)
Then CurrentNode = Backward(T , CurrentNode);
Else CurrentNode = Forward(T , CurrentNode);
End If
End While;
End For;
End.

For instance, let us consider three attributes {a, b, c}, each with three
modalities: {L=low, M=medium, H=high}; the rule tree will be explored by
successively considering premises of rules accordingly to Table 1.

6 From Statistical Implications to Fuzzy Rules

The above algorithm highlights statistical links between conjunctions of fuzzy
attributes and possible fuzzy values for the attribute in conclusion. However,
to apply the modus ponens of fuzzy logics, called generalized modus ponens
(GMP), one must consider fuzzy rules.

6.1 Differences between Fuzzy Rules and Statistical Implications

A fuzzy rule may be considered as the generalization of a classic logical rule;

it associates in its premises a conjunction of fuzzy propositions with, as con-
clusion, a fuzzy proposition. The considered propositions correspond to the
fuzzy attributes of a same item of a database. If X1, X2, . . . , Xn and Y are
fuzzy attributes, having associated respective modalities A1, A2, . . . , An and
B, a fuzzy rule has the form: If “X1 is A1” and “X2 is A2” . . . and “Xn is
An” then “Y is B”, or, more formally “X1 is A1” ∧ “X2 is A2” . . . ∧ “Xn is
An” ⇒ “Y is B”, where A1, A2, . . . , An and B are fuzzy subsets instead of
classical subsets.
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 491

(1) (2) (3)

a=L a=M a=H
a=L∧b=L a=M ∧b=L a=H ∧b=L
a=L∧b=L∧c=L a=M ∧b=L∧c=L a=H ∧b=L∧c=L
a=L∧b=L∧c=M a=M ∧b=L∧c=M a=H ∧b=L∧c=M
a=L∧b=L∧c=H a=M ∧b=L∧c=H a=H ∧b=L∧c=H
a=L∧b=M a=M ∧b=M a=H ∧b=M
a=L∧b=M ∧c=L a=M ∧b=M ∧c=L a=H ∧b=M ∧c=L
a=L∧b=M ∧c=M a=M ∧b=M ∧c=M a=H ∧b=M ∧c=M
a=L∧b=M ∧c=H a=M ∧b=M ∧c=H a=H ∧b=M ∧c=H
a=L∧b=H a=M ∧b=H a=H ∧b=H
a=L∧b=H ∧c=L a=M ∧b=H ∧c=L a=H ∧b=H ∧c=L
a=L∧b=H ∧c=M a=M ∧b=H ∧c=M a=H ∧b=H ∧c=M
a=L∧b=H ∧c=H a=M ∧b=H ∧c=H a=H ∧b=H ∧c=H
(4) (5) (6)
b=L b=M b=H
b=L∧c=L b=M ∧c=L b=H ∧c=L
b=L∧c=M b=M ∧c=M b=H ∧c=M
b=L∧c=H b=M ∧c=H b=H ∧c=H
(7) (8) (9)
c=L c=M c=H
Table 1. The sequence of premises tested by our algorithm

For example, one can write rules such as

“If the pressure is high then the weather is cold” or
“If temperature is low and humidity is high then saturation is near”.
A statistical implication expresses that, when premises are true, it is very
likely its conclusion is true. In a different way, a fuzzy rule allows the evalu-
ation of the fuzzy set associated with its conclusion, given the truth degree
of its premises. The conclusion is then a set of possible values for one fuzzy
attribute, some values being partially appropriate. The fuzzy rule aggregates
its possible counter-examples by extending the range of the possible values of
the attribute in conclusion and by giving them a small degree of membership
to the associated fuzzy set.
To completely define a fuzzy rule, one needs to specify the fuzzy operators
to which it refers. Generally, only one conjunction and one implication are
sufficient, but one often needs one complement and one disjunction as well.
Unfortunately, the number of possible fuzzy operators is infinite, but generally
one may restrict them to some standard sets.

6.2 Main Sets of Fuzzy Operators

With the indexes we have chosen only a basic fuzzy conjunction (“and” oper-
ator) is needed, but if the extracted rules are to be processed by a knowledge
492 M. Bernardet

based system, a fuzzy implication is also necessary, often with a fuzzy disjunc-
tion (“or”) and a fuzzy complement (“not”). Fuzzy logics offer a great choice
of logical operators [21]. Let us summarize the main fuzzy operators.
When there is no ambiguity, we will simplify our notation µ(a) into a,
representing both the proposition a and its truth value.

•A fuzzy complement (fuzzy negation) is generally realized with the stan-

dard complement C(a) = 1 − a; we have only considered this possibility.

•A fuzzy conjunction (fuzzy “and”) must be defined by a t-norm (triangular

norm) >, which is a function from [0, 1] × [0, 1] into [0, 1] characterized by:
>(0, 0) = >(0, 1) = >(1, 0) = 0, >(1, 1) = 1,
>(a, b) = >(b, a) (commutativity),
>(a, >(b, c)) = >(>(a, b), c) (associativity),
∀a, a0 , b, b0 a ≤ a0 and b ≤ b0 ⇒ >(a, b) ≤ >(a0 , b0 ) (monotony).
The axioms on the first three lines keep the properties of the classical
conjunction for classical sets. Other axioms are often added, such as the con-
tinuity of >(x, y) and/or its under-idempotence: >(a, a) ≤ a. One frequently
uses:
. Zadeh’s minimum: >(a, b) = min(a, b),
. the probabilistic intersection: >(a, b) = a × b or
. Lukasiewicz’s bounded (or bold) difference: >(a, b) = max(0, a+b−1).

•A fuzzy disjunction (fuzzy “or”) must be defined by a t-conorm (triangu-

lar conorm) ⊥, which is a function from [0, 1] × [0, 1] into [0, 1], characterized
by:
⊥(1, 1) = ⊥(0, 1) = ⊥(1, 0) = 1, ⊥(0, 0) = 0,
⊥(a, b) = ⊥(b, a) (commutativity),
⊥(a, ⊥(b, c)) = ⊥(⊥(a, b), c) (associativity),
∀a, a0 , b, b0 a ≤ a0 and b ≤ b0 ⇒ ⊥(a, b) ≤ ⊥(a0 , b0 ) (monotony).
The first three lines keep the properties of the classical disjunction. Some
other axioms are often added, such as the continuity of ⊥(x, y) and/or its
over-idempotence: ⊥(a, a) ≥ a. One frequently uses:
. Zadeh’s standard union (or Gödels’s t-norm): ⊥(a, b) = max(a, b),
. the probabilistic union: ⊥(a, b) = a + b − a × b or
. the bounded sum (or Lukasiewicz’s): ⊥(a, b) = min(1, a + b).
The conjunction and the disjunction should be bound by de Morgan’s
laws:
C(>(a, b)) = ⊥(C(a), C(b)) and C(⊥(a, b)) = >(C(a), C(b)).
In this case, the t-norm > and the t-conorm ⊥ are said to be dual relatively
to the fuzzy complement C; amongst t-norms and t-conorms dual relatively
to the standard fuzzy complement C(a) = 1 − a, one may choose
. the minimum and maximum:
>(a, b) = min(a, b), ⊥(a, b) = max(a, b),
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 493

. the probabilistic product and sum:

>(a, b) = a × b, ⊥(a, b) = a + b − a × b or
. the bounded difference and sum:
>(a, b) = max(0, a + b − 1), ⊥(a, b) = min(1, a + b).

•A fuzzy implication is a fonction I from [0, 1] × [0, 1] into [0, 1] which

defines for all truth values of two fuzzy propositions a and b, the truth value
I(a, b) of “if a then b”. This function I may be defined in different ways in
fuzzy logics, which are equivalent in classical logics.
1) In classical logics one may define I by:
I(a, b) = ¬ a ∨ b, (15)
which becomes in fuzzy logics:
I(a, b) = ⊥(C(a), b) (16)
2) Classical logics allow also to define I by:
I(a, b) = max{x ∈ {0, 1} | (a ∧ x) ≤ b}, (17)
which becomes in fuzzy logics:
I(a, b) = sup{x ∈ [0, 1] | >(a, x) ≤ b} (18)
3) Formula (15) may also be written
I(a, b) = ¬ a ∨ (a ∧ b), (19)
or
I(a, b) = (¬ a ∧ ¬ b) ∨ b; (20)
these formulas become in fuzzy logics:
I(a, b) = ⊥(C(a), >(a, b)), (21)
or
I(a, b) = ⊥(>(C(a), C(b)), b), (22)
where >, ⊥ and C must satisfy de Morgan’s law.
Definitions (15), (17) (19) and (20) are equivalent, which is not true for
definitions (16), (18), (21) and (22), which consider fuzzy truth values instead
of classical truth values 0 or 1; these 4 last formulas allow definitions of several
classes of fuzzy implications.
- S-implications are defined using formula (16), which specifies a fuzzy
implication I given a t-conorm ⊥: I(a, b) = ⊥(C(a), b). One may then define:
- for the maximum (dual intersection: the minimum),
Kleene-Dienes’ implication: I(a, b) = max(1 − a, b);
- for the probabilistic union (dual intersection: the probabilistic product),
Reichenbach’s implication: I(a, b) = 1 − a + a × b;
- for the bounded sum (dual intersection: the bounded difference),
Lukasiewicz’s implication: I(a, b) = min(1, 1 − a + b).
- R-implications are defined by formula (21) which specifies one implication
given one t-norm >: I(a, b) = sup{x ∈ [0, 1] |>(a, x) ≤ b}. This allows to
define:
494 M. Bernardet

- for the minimum, Gödel’s implication:

I(a, b) = 1 if a ≤ b oror b if a > b.
- for the probabilistic product, Goguen’s implication:
I(a, b) = 1 if a ≤ b oror b/a if a > b.
- for the bounded difference, Lukasiewicz’s implication:
I(a, b) = min(1, 1 − a + b).
- QL-implications use relation (22) with one t-norm > and its dual t-conorm
⊥. This class of implications did not prove interesting within our studies, so
we will not go into it.

Our algorithms also need one aggregation operator , which may be defined
in numerous ways. However, since we want an averaging evaluation of the
implication and since we need a mechanism to allow exclusion of abnormal
records, we have chosen the arithmetic mean, which allows the use of standard
deviations:
n
Aggregation(µ1 (x, y), . . . , µn (x, y)) = 1/n
X
µi (x, y). (23)
i=1

7 Fuzzy Operators to be Used with Generalized Modus

Ponens
When the extracted rules are to be used in a knowledge-based system by
application of the generalized modus ponens (GMP), the choice of fuzzy op-
erators must be coherent with the operators chosen then. Let us recall that
the GMP is the following inference scheme:
If X is A then Y is B
X is A’

Y is B’

For one implication µa⇒b (µa (x), µb (y)) = I(µa (x), µb (y)) and one t-norm >,
one can write:
µb0 (y) = sup {>(µa0 (x), I(µa (x), µb (y) )} (24)
x∈A0

7.1 Our Method

To choose a set of fuzzy operators, we have defined a specific index which

evaluates fuzzy rules given this set of operators [4]. This index, called the
GMP-pertinence, is determined by applying the GMP on the considered rule
for a random sample of a database. If the distance between the truth-values
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 495

of the inferred conclusions and the observation is below a threshold chosen

by the operator, the example is added to the set of records for which the test
is correct, otherwise it is added to the set of records for which the test fails.
We define the GMP-pertinence of the rule PP given aPset
P of fuzzy operators as
its rate of good examples: GM P pert = n+ / ( n+ + n− ).
x y x y
A comparative study of fuzzy implication operators [20] has shown that
the GMP gives good results by combining the minimum and bold intersec-
tions with Lukasiewicz’s, Kleene-Dienes’ and Gödel-Brouwer’s implications.
We have added to these operators the probabilistic conjunction and Goguen’s
implication because of their probabilistic nature. So, we have limited our trials
to combinations of one of these t-norms with one of these implications. To il-
lustrate our method, we first present a simplified example, then we summarize
the results obtained on more consequent databases.

7.2 One Simple Example

Let us consider two attributes “Size” and “Shoe size” in a set of individuals
who can be grouped in two fuzzy classes: the small persons with small size
feet, and the tall persons with large size feet. To these classes correspond 2
rules:
“small size” ⇒ “small shoe size”
“tall size” ⇒ “large shoe size”

First Benchmark
In a first benchmark, the items that do not belong to any class and which
may be considered as noisy data make up 5% of all data. We have studied the
GMP-pertinences for the 4 possible rules associating one fuzzy modality of
the first attribute (size) to one modality of the second attribute (shoe size).
Figure 4 describes the data set, in which points outside the ellipses represents
noisy data.

A) For the rule “size=small” ⇒ “shoe size=small”

The GMP-pertinences of the fuzzy operators calculated with various t-norms
are then rather close: 0.866 for bold intersection, 0.850 for probabilistic in-
tersection and 0.844 for Zadeh’s intersection. This rule is interesting because
such levels of confidence indicate that about 85% of the persons of small size
have a small shoe size. Table 7.2 shows the GMP-pertinence for sets of oper-
ators combining each considered implication with every t-norm. Results are
rather close whichever t-norm or implication is used.

B) For the rule “size=tall” ⇒ “shoe size=small”

Confidence levels calculated now with different t-norms are 0.148 for bold
intersection, 0.164 for probabilistic intersection and 0.170 for Zadeh’s inter-
section. This rule is not interesting since it is verified for only about 16% of
tall persons.
496 M. Bernardet

Fig. 4. Data distribution in the first benchmark

Implication bold conjunction Zadeh’s conjunction probabilistic conjunction

Gödel-Brouwer 0,895923 0,900154 0,897859
Goguen 0,899513 0,899546 0,900154
Kleene-Dienes 0,887846 0,897010 0,890855
Lukasiewicz 0,900154 0,896241 0,897908

Table 2. GMP-pertinence of the fuzzy operators for

the rule “size=small” ⇒ “shoe size=small”

C) For the rule “size=tall” ⇒ “shoe size=large”

Confidence levels with different t-norms are 0.780 for bold intersection, 0.760
for probabilistic intersection and 0.754 for Zadeh’s. This rule is interesting with
a conditional probability about 76.5%. Table 7.2 below shows rather significant
differences (1 to 5%) on the quality of the operators for each t-norm. With bold
t-norm, three implications, Lukasiewicz’s, Goguen’s and Gödel-Brouwer’s give
good results. With the probabilistic t-norm, implications of Goguen, Gödel-
Brouwer and Lukasiewicz give good and rather close results, since their quality
is above 92%.

Implication bold conjunction Zadeh’s conjunction probabilistic conjunction

Gödel-Brouwer 0,921295 0,926686 0,923805
Goguen 0,92604 0,926031 0,926686
Kleene-Dienes 0,876628 0,919381 0,889767
Lukasiewicz 0,926686 0,918095 0,921885

Table 3. GMP-pertinence of the fuzzy operators for

the rule “size=tall” ⇒ “shoe size=large”
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 497

Second Benchmark
In a second benchmark (Fig. 5) we have increased the rate of noisy data,
which accounts now for 37% of all data.

Fig. 5. Data distribution in the 2nd benchmark

A) For the rule “size=small” ⇒ “shoe size=small”

Confidence levels calculated with various t-norms are rather near: 0.523 for
bold intersection, 0.510 for probabilistic intersection and 0.516 for Zadeh’s
intersection. This sample is critical with a confidence about 52%, which jus-
tifies its interest although results of this rule may be erroneous. This interest
is confirmed by a support near 28%.
Table 7.2 shows that qualities of operators are rather low (less than 85%).
Again, some implications may be grouped accordingly with the t-norms, par-
ticularly the implications of Gödel-Brouwer and Goguen on the one hand, and
those of Lukasiewicz and Goguen on the other hand.

Implication bold conjunction Zadeh’s conjunction probabilistic conjunction

Gödel-Brouwer 0,800913 0,848043 0,825563
Goguen 0,834291 0,840908 0,848043
Kleene-Dienes 0,766130 0,797044 0,778592
Lukasiewicz 0,848043 0,781826 0,808287

Table 4. GMP-pertinence of the fuzzy operators for

the rule “size=small” ⇒ “shoe size=small”

B) For the rule “size=tall” ⇒ “shoe size=large”

498 M. Bernardet

Confidence levels calculated with different t-norms are rather close: 0.726 for
bold intersection, 0.700 for probabilistic intersection and 0.688 for Zadeh’s.
So, this rule is interesting with a conditional probability about 70%.

Implication bold conjunction Zadeh’s conjunction probabilistic conjunction

Gödel-Brouwer 0,915288 0,936121 0,925429
Goguen 0,930192 0,930027 0,936121
Kleene-Dienes 0,872485 0,906273 0,882153
Lukasiewicz 0,936121 0,897394 0,91496

Table 5. GMP-pertinence of the fuzzy operators for

the rule “size=tall” ⇒ “shoe size=large”

Table 7.2 shows rather significant differences (1 to 5%) on the evaluation

of the operators for each t-norm. One can remark that with bold t-norm,
three implications, Lukasiewicz’s, Goguen’s and Gödel-Brouwer’s, give good
results. With the probabilistic t-norm, implications of Goguen, Gödel-Brouwer
and Lukasiewicz give good and rather close results: their GMP-pertinence is
above 92%.
Results of Table 7.2 are similar to those of 7.2, with a better GMP-
pertinence. One can again associate the same operators as with table 3, but
with much higher scores, which confirms these associations.

7.3 Results on real databases

We have tried out our algorithms on several databases found in the UCI repos-
itory [19], in particular “Wisconsin Breast Cancer Database”, which consists
of 699 items with 10 attributes and two classes, “Wine Recognition Database”
with 178 items, 13 attributes and three classes and “Ionosphere Database”
with 351 items, 39 attributes and 2 classes. The results are similar to those
highlighted by our previous example, but the differences on GMP-pertinences
of operators are less strong than in our second benchmark, for which the pro-
portion of noisy data had been deliberately strengthened. Let us consider a
few examples extracted from our results.
For “Wisconsin Breast Cancer Database”, with 3 fuzzy partitions on each
attribute, a minimal confidence of 0.8, a minimal intensity of implication of
0.9, a support of 5% and at most 3 propositions in the premises, we get 331
rules; if one pushes the search to 6 propositions, we obtain 814 rules. Going
to 9 propositions brings few supplementary rules with a total of 870.
Considering the evolution of the rule numbers according to the maximum
number of premises and the number of classes (Table 7.3), one remarks that
the number of rules decreases when the number of classes increases, until 8
classes.
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 499

at most 3 premises at most 6 premises at most 9 premises

2 classes 407 rules 908 rules 954 rules
3 classes 331 rules 814 rules 870 rules
4 classes 250 rules 632 rules 678 rules
5 classes 267 rules 464 rules 466 rules
6 classes 253 rules 471 rules 474 rules
7 classes 255 rules 432 rules 434 rules
8 classes 242 rules 410 rules 412 rules
9 classes 395 rules 672 rules 674 rules

Table 6. Evolution of the number of rules with the number of classes

for “Wisconsin Breast Cancer Database”

The profusion of rules with small numbers of classes is offset by the im-
precision of the rules: the average confidence of rules with 2 classes is much
weaker than that obtained with more classes. With 2 classes and at most 6
premises, only 35 rules (7%) have a confidence of 1, while with 9 classes and
6 premises, 388 rules (57%) have the same confidence of 1. Increasing the
number of premises beyond 6 brings little improvement because addition of
attributes to rules only specializes the rules with a confidence under 1.
For example, the rule If Clump Thickness is “very small” then Class is
“benign”, appears with a confidence of 0.964, a support of 28% and a GMP-
pertinence of 91.1%, 90.3% or 89.6% depending on the fuzzy operators. Spe-
cialization of this rule by adding supplementary attributes increases confidence
by reducing the support, until it reaches a confidence of 1; the rule is then: If
Clump Thickness is “very small” and Single Epithelial Cell Size is “very small”
and Bare Nucleii=“very little” then Class is “benign”. Its support is then 26%
with a GMP-pertinence of 1.
Results on other databases are similar, but, due to the highest number of
attributes, the numbers of generated rules are much larger. Using the same
thresholds and 3 classes by attribute, one obtains for “Wine Data Base” 1092
rules of at most 3 premises and 13470 of at most 6 premises. With “Ionosphere
Data Base” one extracts 13824 rules with at most 3 premises. A more severe
choice of thresholds is then needed to reduce these high numbers of rules.
Let us consider another example, from the results on “Wine Data Base”.
The rule If Magnesium is “very little” then Wine is “type2” appears with a
confidence of 0.894, and when specializing it by adding one attribute, we
get 3 new rules with a confidence of 1, such as the rule If Magnesium is
“very little” and Flavanoids is “little” then Wine is “type2”. Adding one more
attribute gives 9 more rules with a full confidence of 1. Comparisons between
qualities of operators using GMP-pertinences confirm again our conclusions
on the choice of operators, the association between Lukasiewicz’ implication
and Lukasiewicz’ bounded sum and difference appearing slightly better.
500 M. Bernardet

8 Synthesis of these studies

We have remarked on all examples that, whatever the rule, for a given t-norm,
the same implications always have the best GMP-pertinence in the evalua-
tion of the GMP. For the bold t-norm, one can group Goguen’s implication
and Lukasiewicz’s, these two implications being the best. For Zadeh’s t-norm,
implications of Gödel-Brouwer and Goguen are the best. For probabilistic t-
norm, one can group implications of Gödel-Brouwer and Goguen, as the best
ones. To summarize these results, the sets of operators that appear then as
the best with the GMP associate:
- Lukasiewicz’s t-norm and Lukasiewicz’s implication,
>(a, b) = max(0, a + b − 1)
I(a, b) = min(1, 1 − a + b).
- Gödel’s t-norm (Zadeh’s minimum) and Gödel-Brouwer’s implication,
>(a, b) = min(a, b)
I(a, b) = 1 if a ≤ b or b if a > b.
- Probabilistic t-norm and Goguen’s implication.
>(a, b) = a × b
I(a, b) = 1 if a ≤ b or b/a if a > b.
Thus, the sets of operators that appear experimentally the best to use with
the GMP, are those that associate a t-norm > with the R-implication I that
it defines:

I(a, b) = sup{x ∈ [0, 1] | >(a, x) ≤ b} (25)

This result is justified by theory, in particular by [16], which shows that
the best implication to apply with the GMP for a given t-norm is the residue
of this t-norm (the definition of an R-implication is indeed the residue of
the associated t-norm). The proof lies on the facts that for a given truth
value a of the premises, a given truth value function I(a, b) of the implication
between a and the conclusions b, the GMP should be non decreasing between
its arguments, since the truer the antecedent and the truer the implication,
the truer the consequent should be. Moreover, since the neutral element of the
GMP should be 0, and its unit 1, the GMP should be realized with a t-norm.
In order to have the most powerful rules, one has to choose I(a, b) as large as
possible, giving then definition (25).

9 Operators for Fuzzy Rule Reduction

When rules are not extracted to build knowledge based systems, but to give
human experts a synthetic view of the database, the number of rules extracted
is often too high to be studied, and methods to reduce sets of fuzzy rules are
welcome. For this kind of applications of knowledge discovery, we have studied
methods allowing aggregation of rules.
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 501

9.1 A First Scheme of Rule Reduction

In classical logics, one may write (a ⇒ c) ∧ (b ⇒ c) ` (a ∨ b) ⇒ c, and if one

wants to process similarly with fuzzy rules without having to reevaluate the
inferred fuzzy rule (a ∨ b) ⇒ c, it should be rigorously correct to write
∀(x, y) ∈ [0, 1] × [0, 1] : µ(a⇒c)∧(b⇒c) (x, y) = µ((a∨b)⇒c (x, y),
or, with a t-norm > and its complementary t-co-norm ⊥:
∀(x, y) ∈ [0, 1] × [0, 1] : µa⇒c (x, y) > µb⇒c (x, y) = µ(a∨b)⇒c (x, y)
If we note α = µa (x, y), β = µb (x, y), γ = µc (x, y) and I(α, β) =
µa⇒c (x, y), we had to find which, if any, of the operator sets allows to write
the condensed form:
I(α, γ) > I(β, γ) ≡ I(α ⊥ β, γ) or
>(I(α, γ), I(β, γ)) ≡ I(⊥(α, β), γ)) (26)
We have then considered four fuzzy implications:
- Kleene-Dienes’ implication: IKD (a, b) = max(1 − a, b);
- Gödel-Brouwer’s implication: IGB (a, b) = 1 if a ≤ b or b if a > b;
- Goguen’s implication: IGog (a, b) = 1 if a ≤ b or b/a if a > b;
- Lukasiewicz’s implication: IL (a, b) = min(1, 1 − a + b).
with 3 dual pairs of t-norm and t-conorm:
- Zadeh’s minimum and maximum:
>Z (a, b) = min(a, b), ⊥Z (a, b) = max(a, b),
- the probabilistic product and sum:
>Pr (a, b) = a × b, ⊥Pr (a, b) = a + b − a × b,
- Lukasiewicz’s difference and sum:
>L (a, b) = max(0, a + b − 1), ⊥L (a, b) = min(1, a + b).
We have then proved that with Zadeh’s t-norm and t-conorm, all the im-
plications we considered verify relation (26), but that none of these implica-
tions combined with the probabilistic t-norm/t-conorm or with Lukasiewicz’s
bounded sum and difference verify it.

Proof
A) With Zadeh’s >Z (a, b) = min(a, b) and ⊥Z (a, b) = max(a, b), one may take
into account the fact that fuzzy implications are monotonously decreasing with
their first argument: the truth-value of any fuzzy implications must increase
as the truth of its antecedent decreases. So, for any fuzzy implication I
- when α ≤ β
I(α, γ) ≥ I(β, γ), so >Z (I(α, γ), I(β, γ)) = min(I(α, γ), (β, γ)) = I(β, γ),
while I(⊥Z (α, β), γ) = I(max(α, β), γ) = I(β, γ).
Therefore, in this case, >Z (I(α, γ), I(β, γ)) = I(β, γ) = I(⊥Z (α, β), γ), and
by symmetry on α and β, this result is always true:
>Z (I(α, γ), I(β, γ)) ≡ I(⊥Z (α, β), γ).

B) With the probabilistic t-norm and t-conorm, let us consider counter-

examples:
502 M. Bernardet

- for Kleene-Dienes’ implication, with α = 0.5, β = 0.5, γ = 0.6,

>Pr (IKD (α, γ), IKD (β, γ)) = 0.3, but IKD (⊥Pr (α, β), γ) = 0.6;
- for Lukasiewicz’s implication, with α = 0.5, β = 0.7, γ = 0.5,
>Pr (IL (α, γ), IL (β, γ)) = 0.8, but IL (⊥Pr (α, β), γ) = 0.65;
- for Gödel-Brouwer’s implication with α = 0.5, β = 0.5, γ = 0.6,
>Pr (IGB (α, γ), IGB (β, γ)) = 1, but IGB (⊥Pr (α, β), γ) = 0.6;
- for Goguen’s implication with α= 0.5, β=0.5, γ= 0.6
>Pr (IGog (α, γ), IGog (β, γ)) = 1, but IGog (⊥Pr (α, β), γ) = 0.8.

C) With Lukasiewicz’s difference and sum, let us consider counter-examples:

- for Kleene-Dienes’ implication, with α = 0.5, β = 0.5, γ = 0.6,
>L (IKD (α, γ), IKD (β, γ)) = 0.1, but IKD (⊥L (α, β), γ) = 0.6;
- for Lukasiewicz’s implication, with α = 0.5, β =0.7, γ = 0.5,
>L (IL (α, γ), IL (β, γ)) = 0.8, but IL (⊥L (α, β), γ) = 0.5;
- for Gödel-Brouwer’s implication and α = 0.5, β = 0.5, γ = 0.6,
>L (IGB (α, γ), IGB (β, γ)) = 1, but IGB (⊥L (α, β), γ) = 0.6;
- for Goguen’s implication and α = 0.5, β = 0.5, γ = 0.6,
>L (IGog (α, γ), IGog (β, γ)) = 1, but IGog (⊥L (α, β), γ) = 0.6.
So, whatever the implication, among the considered pairs of dual t-norms
and t-conorms, only Zadeh’s minimum and maximum verifies relation (26),
which allows easy fuzzy rule reduction.

9.2 A Second Scheme of Rule Reduction

Similarly, to keep the classical reduction scheme (a ⇒ b) ∧ (a ⇒ c) ` a ⇒

(b ∧ c) or, with the same condensed notations as above
I(α, β)>I(α, γ)) ≡ I(α, β>γ) or
>(I(α, β), I(α, γ)) ≡ I(α, >(β, γ)), (27)
we have proved that with Zadeh’s t-norm and t-conorm, all the considered
implications verify relation (27), but that none of these implications combined
with probabilistic t-norm or with Lukasiewicz’s t-norm verify it.

Proof
A) With Zadeh’s >Z (a, b) = min(a, b) and ⊥Z (a, b) = max(a, b), one may take
into account the fact that fuzzy implications are monotonously increasing with
their second argument: the truth-value of any fuzzy implication must increase
when the truth of its conclusion increases. So, for any fuzzy implication I
- when β ≤ γ
I(α, β) ≤ I(α, γ), so >Z (I(αβ), (α, γ)) = min(I(α, β), I(α, γ)) = I(α, β),
while I(α, >Z (β, γ)) = I(α, min(β, γ)) = I(α, β).
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 503

Therefore, in this case, >Z (I(α, β), I(α, γ)) = I(α, β) = I(α, >Z (β, γ)), and
by symmetry on β and γ, this result is always true:
>Z (I(α, β), I(α, γ)) ≡ I(α, >Z (β, γ)).

B) With probabilistic t-norm and t-conorm, let us consider counter-examples:

- for Kleene-Dienes’ implication and α = 0.4, β = 0.8, γ = 0.5,
>Pr (IKD (α, b), IKD (α, γ)) = 0.48, but IKD (α, >Pr (β, γ)) = 0.6;
- for Lukasiewicz’s implication and α = 0.6, β = 0.3, γ = 0.5,
>Pr (IL (α, b), IL (α, γ)) = 0.63, but IL (α, >Pr (β, γ) = 0.4;
- for Gödel-Brouwer’s implication and α = 0.6, β = 0.7, γ = 0.8,
>Pr (IGB (α, β), IGB (α, γ)) = 1, but IGB (α, >Z (β, γ) =0.56;
- for Goguen’s implication and α = 0.6, β = 0.6, γ = 0.6,
>Pr (IGog (α, β), IGog (α, γ) = 1, but IGog (α, >Pr (β, γ) = 0.6.

C) With Lukasiewicz’s t-norm and t-conorm, let us consider counter-examples:

- for Kleene-Dienes’ implication and α = 0.4, β = 0.6, γ = 0.7,
>L (IKD (α, β), IKD (α, γ) = 0.3, but IKD (α, >L (β, γ)) = 0.6;
- for Lukasiewicz’s implication and α = 0.6, β = 0.3, γ = 0.6,
>L (IL (α, β), IL (α, γ)) = 0.7, but IL (α, >L (β, γ) = 0.5;
- for Gödel-Brouwer’s implication and α = 0.5, β = 0.6, γ = 0.6,
>L (IGB (α, β), IGB (α, γ)) = 1, but IGB (α, >L (β, γ) = 0.2;
- for Goguen’s implication and α = 0.5, β =0.6, γ= 0.6:
>L (IGog (αβ), IGog (αγ)) = 1, but IGog (α, >L (β, γ) = 0.4.
So, whatever the implication, among the considered t-norms only Zadeh’s
minimum verifies relation (27), which allows easy fuzzy rule reduction.

One may remark here that among the sets of fuzzy operators that appear
as the best with the GMP, we have found Gödel’s t-norm (Zadeh’s minimum)
and Gödel-Brouwer’s implication. Therefore, this set of fuzzy operators may
be considered as the most interesting when one wants to extract rules for a
knowledge based system and also to reduce the extracted rules within a same
application. These results also illustrate that fuzzy rules cannot generally be
treated as classical rules.

10 Conclusion
We have described a generalization of statistical implication indexes to fuzzy
knowledge discovery. The first operation needed to compute these indexes is
the choice of fuzzy partitions to convert numerical or symbolic attributes into
fuzzy ones. We have justified our choice of three classical statistic indexes:
the support, the confidence and the less common, but powerful, intensity of
implication. We have then explained how we have adapted these indexes to
504 M. Bernardet

fuzzy attributes by replacing cardinalities of crisp sets by cardinalities of fuzzy

sets, then we have described one algorithm to explore the set of possible rules.
The indexes we use to extract fuzzy rules highlight statistical links be-
tween conjunctions of fuzzy attributes and fuzzy conclusions, but they do not
evaluate the associated fuzzy rules, which depend on the chosen fuzzy opera-
tors. Since fuzzy operators are numerous, we have evaluated sets of standard
operators by applying the generalized modus ponens (GMP) on the database
items and by comparing its results to the effective conclusions. After a simpli-
fied example illustrating these mechanisms, we have given a summary of our
results on more consequent databases.
We have then observed that the sets of fuzzy operators which give the best
results with the GMP, associate a t-norm and the related R-implication:
- Lukasiewicz’s t-norm (bold t-norm) and Lukasiewicz’s implication,
- Gödel’s t-norm (Zadeh’s minimum) and Brouwer-Gödel’s implication,
- probabilistic t-norm and Goguen’s implication.
To allow an easy reduction of the number of rules proposed to human
experts, we have studied methods to cluster fuzzy rules. We have then shown
that among the considered sets of operators, using Zadeh’s minimum as t-norm
and maximum as t-conorm is the best choice, independently of the implication.
Since Gödel-Brouwer’s implication gives the best results with the GMP when
one uses Zadeh’s minimum and maximum, this implication seems a rather
good compromise. Associating Lukasiewicz’s implication with Lukasiewicz’s
t-norm and t-conorm shows better GMP-pertinence, but does not allow clas-
sical schemes of rule reduction.
Finally, we must remark that the increase of computing complexity in-
duced by the use of fuzzy logics is relatively small for rule extraction, since
instead of increasing by one (integer number) the counters of good and bad
examples, fuzzy logics add membership degrees (real numbers). The opera-
tions of fuzzification, the choice of fuzzy operators and the reductions of rules
are more complex, but the advantages of using fuzzy logics may compensate
for this, because intervals on continuous attributes are expressed by more
expressive fuzzy labels and since abrupt threshold are avoided.

References
1. R. Agrawal, T. Imielinski, and A. N. Swami. Mining association rules between
sets of items in large databases. In Peter Buneman and Sushil Jajodia, editors,
Proceedings of the 1993 ACM SIGMOD International Conference on Manage-
ment of Data, pages 207–216, Washington, D.C., 1993.
2. J. Aguilar-Martin and R. Lopez De Mantaras. The process of classification and
learning the meaning of linguistic descriptors of concepts. In M. M. Gupta and
E. Sanchez, editors, Approximate reasoning in decision analysis, pages 165–175.
North Holland, 1982.
3. M. Bernadet. Basis of a fuzzy knowledge discovery system. In Conf. PKDD’2000
- LNAI 1910, pages 24–33. Springer-Verlag, 2000.
 Fuzzy Knowledge Discovery Based on Statistical Implication Indexes 505

4. M. Bernadet. A comparison of operators in fuzzy knowledge discovery. In Conf.

IPMU’2004, volume 2, pages 731–738, July 2004.
5. M. Bernadet. A study of data partitioning methods for fuzzy knowledge discov-
ery. In Conf. IPMU’2006, pages 1396–1402, July 2006.
6. M. Bernadet, G. Rose, and H. Briand. Fiable and fuzzy fiable: two learn-
ing mechanisms based on a probabilistic evaluation of implications. In
Conf. IPMU’96 (Information Processing and Management of Uncertainty in
Knowledge-Based Systems), volume 2, pages 911–916, July 1996.
7. J. C. Bezdek and J. D. Harrisand. Fuzzy partitions and relations: An axiomatic
basis for clustering. Fuzzy Sets and Systems, 1:111–127, 1978.
8. A. Bonarini. Evolutionary learning of fuzzy rules: competition and cooperation.
In Fuzzy Modeling: Paradigms and Practice, pages 265–284. Kluwer Academic
Press, 1996.
9. H. Briand, L. Fleury, R. Gras, Y. Masson Y. J., and Philippe. A statistical
measure of rules strength for machine learning. In WOCFAI 1995, pages 51–62,
1995.
10. J. J. Buckley and K. Hayashi. Neural networks for fuzzy systems. Fuzzy Sets
and Systems, pages 265–276, 1995.
11. M. Delgado, D. Sánchez, and M. A. Vila. Fuzzy cardinality based evaluation of
quantified sentences. Int. Journal of Approximate Reasoning, 23:23–66, 2000.
12. L. Fleury and Y. Masson. Intensity of implication: a measurement in machine
learning. In IEA/AIE’95, pages 621–629, June 1995.
13. W. J. Frawley, G. Piatetsky-Shapiro, and C. J. Matheus. Knowledge discovery
in databases: An overview. AI Magazine, 13:57–70, 1992.
14. R. Gras, R. Couturier, F. Guillet, and F. Spagnolo. Extraction de règles en
incertain par la méthode implicative, extraction des connaissances : Etat et
perspectives. In RNTI-E-5, pages 385–389. Cepaduès Editions, 2006.
15. R. Gras and A. Larher. L’implication statistique, une nouvelle méthode
d’analyse de données. Mathématiques, Informatique et Sciences Humaines,
120:5–31, 1992.
16. P. Hajek. Metamathematics of Fuzzy Logic. Kluwer Academic, 1998.
17. S. K. Halgamuge and M. Glesner. Neural networks in designing fuzzy systems
for real world applications. Fuzzy Sets and Systems, 65:1–12, 1994.
18. A. Heinz. Adaptive fuzzy neural trees. In IDA-95 Symposium, pages 70–74,
August 1995.
19. S. Hettich and S. D. Bay. The UCI KDD Archive. Univ. of California, Irvine, De-
partment of Information and Computer Science, 1999. [http://kdd.ics.uci.edu].
20. E. E. Kerre. A comparative study of the behavior of some popular fuzzy impli-
cations on the generalized modus ponens. In Fuzzy Logic for the management
of uncertainty, pages 281–295. John Wiley & Sons, Inc., 1992.
21. G. J. Klir and Bo Yuan. Fuzzy Sets and Fuzzy Logics - Theory and Applications.
PrenticeHall. Englewood Cliffs, 1995.
22. L. Lesmo, L. Saitta, and P. Torasso. Fuzzy production rules: a learning method-
ology. In Advances in Fuzzy Sets, Possibility Theory and Applications, pages
181–198. Plenum Press, 1983.
23. T. Murata, H. Ishibuchi, H. Nakashima, and M. Gen. Fuzzy partition and
input selection by genetic algorithms for designing fuzzy rule-based classification
systems. In LNCS 1447, pages 407–416. Springer-Verlag, 1998.
24. A. Ralescu. Cardinality, quantifiers and the aggregation of fuzzy criteria. Fuzzy
Sets and Systems, 69:355–365, 1995.
506 M. Bernardet

25. J. Rives. FID3: Fuzzy induction decision tree. In ISUMA’90, pages 457–462,
December 1990.
26. J. A. Roubos, M. Setnes, and J. Abonyi. Learning fuzzy classification rules from
data. In Developments in Soft Computing, pages 108–115. Springer-Verlag, 2001.
27. F. Spagnolo and R. Gras. A new approach in Zadeh classification: fuzzy im-
plication through statistic implication. In NAFIPS-IEEE 3rd Conference of
the North American Fuzzy Information Processing Society, pages 425–429, June
2004.
28. R. Weber. Fuzzy-ID3: a class of methods for automatic knowledge acquisition.
In 2nd International Conference on Fuzzy Logic and Neural Networks, pages
265–268, July 1992.
29. M. Wygralak. Questions of cardinality of finite fuzzy sets. Fuzzy Sets and
Systems, 102:185–210, 1999.
30. L. A. Zadeh. Probability measures of fuzzy events. Journal of Mathematical
Analysis and Applications, 23:421–427, 1968.
31. L. A. Zadeh. Fuzzy logic and its application to approximate reasoning. Infor-
mation Processing, 74:591–594, 1974.
32. J. Zeidler and M. Schlosser. Continuous valued attributes in fuzzy decision trees.
In Conf. IPMU’96 (Information Processing and Management of Uncertainty in
Knowledge-Based Systems), pages 395–400, 1996.
33. D. A. Zighed, S. Rabaseda, R. Rakotomalala, and F. Feschet. Discretization
methods in supervised learning. In Encyclopedia of Computer Science and Tech-
nology, volume 40, pages 35–50. Marcel Dekker, 1999.
About the editors

Régis Gras is an Emeritus professor at Polytech’Nantes (Polytechnic gradu-

ate School of Nantes University, France) and he is a member of the “KnOwl-
edge and Decision” team (KOD) in the Nantes-Atlantic Laboratory of Com-
puter Sciences (LINA, CNRS UMR 6241), since 1998. He received a PhD
degree of mathematics in 1979 from the University of Rennes, France. He
used to be the Chair of the French International Commission on Teaching
of Mathematics (1995–1998), then a member of the Committee on Teaching
of Mathematics of the European Mathematics Society (1997–2003). He has
designed a set of methods gathered in his “Statistical Implicative Analysis”
original approach. Since, he continues to develop and extend this approach to
data mining issues. He served as PC Chairs of the 4 editions of SIA confer-
ence (France 2001, Sao Paulo Brazil 2003, Palermo Italy 2005, and Castellon
Spain 2007). He is the author of 4 books, and a co-editor of 3 books of chapters.

Einoshin Suzuki received his Bachelor, Master, and Doctor of Engineering

Degrees all from the University of Tokyo in 1988, 1990, and 1993, respectively.
He joined Tokyo Institute of Technology as an assistant professor (1993–1996)
then Yokohama National University as a lecturer professor (1996–1997) and
was promoted to an associate professor (1997–2006). Since April 2006, he is
with Kyushu University as a full professor. He has obtained the Best Paper
Award of the Japanese Society for Artificial Intelligence twice. He has served
as the Honorary Chair of EGC-07, the PC Chair of DS-04, and the PC Vice
Chair of ICDM-04 and is serving as the Steering Committee Chair of the Inter-
national Conference on Discovery Science and a PC Co-Chair of PAKDD-08.

Fabrice Guillet is an associate professor in Computer Science at Poly-

tech’Nantes, and he is a member of the “KnOwledge and Decision” team
(KOD) in the Nantes-Atlantic Laboratory of Computer Sciences (LINA,
CNRS UMR 6241) since 1997. He received a PhD degree in Computer Sci-
ences in 1995 from the Ecole Nationale Superieure des Telecommunications
de Bretagne. He is a founder of the “Knowledge Extraction and Management”
508 M. Bernardet

French-speaking association of research (EGC1 ), and he is also involved in

the steering committee of the annual EGC French-speaking conference since
2001. His research interests include knowledge quality and knowledge visual-
ization in the frameworks of Data Mining and Knowledge Management. He
has recently co-edited with H. Hamilton a refereed book of chapters entitled
“Quality Measures in Data Mining” published by Springer in 2007.

Filippo Spagnolo received his Bachelor degree in 1972 (Università di

Palermo, Italy), and PhD (Didactics of Mathematics) in 1995 (University
of Bordeaux, France). Researcher in Groups of research in University of
Palermo in “Matematiche Complementari” (Mathematics Education, History
of Mathematics and Fundamenta of Mathematics), 1979–2001. Associate Pro-
fessor in “Matematiche Complementari”, since 2004. Editorial Board of many
international reviews: Mediterranean Journal for Research in Mathematics
Education, Canadian Journal of Science Mathematics and Technology Edu-
cation, Acta Didactica Universitatis Comenianae (Mathematics). Editorial in
Chief of review “Quaderni di Ricerca in Didattica (Mathematics)”, G.R.I.M.
since 1990, Palermo, Italy. Co-ordinator of PhD “Storia e Didattica della
Matematica, Storia e Didattica della Fisica, Storia e Didattica della Chim-
ica”, University of Palermo with a consortium of 4 Universities of Italy and 4
Universities in Europe.

About the manuscript coordinator

Bruno PINAUD received his engineer diploma in computer science from

Polytech’Nantes (Polytechnic graduate School of Nantes University, France)
in 2001 and his PhD in 2006. He is currently an adjunct assistant Professor at
Polytech’Nantes and an associate member of the “KnOwledge and Decision”
team (KOD) in the Nantes-Atlantic Laboratory of Computer Sciences (LINA
CNRS 2729). His current main research activities are about graph drawing
with metaheuristics and some applications in data-mining and knowledge vi-
sualization.

1
http://www.polytech.univ-nantes.fr/associationEGC
Index

χ2 , 13 Bayes’ theorem, 167

χ2 distance, 31 Bayesian inference, 163
δ-free descriptions, 473 Bayesian information criteria, 406
behavioral indicators, 300
a posteriori analysis, 112, 321 behavioral profile, 301
a priori, 102 behaviour group, 92
a priori analysis, 112, 248, 321 Beta distribution, 167
a priori matrix, 321 BIC, see Bayesian information criteria
Abdut, 259 binary variables, 13, 43
additional variable, 30 Binomial, 14, 61
Agence Nationale Pour l’Emploi, 305 Binomial distribution, 33, 61
Agence pour l’Emploi des Cadres, 300 bipolar dimensions, 302
aggregation operator, 494 Boolean logics, 481
algebraic context variables, 82 bootstrap procedure, 454
algorithmic technique, 325
androgynous form, 127 CAIMAN matching service, 231
ANOVAF, 213 CART, 400
ANPE, see Agence Nationale Pour cartesian graph of functions, 100
l’Emploi CAS, see Computer Algebra Systems
Anthropological Theory of Didactics, 83 causal conception, 167
APEC, see Agence pour l’Emploi des causal relationships, 13
Cadres causality, 20
Aplusix learning environment, 76 CFA, see Correspondence Factor
Apriori-like algorithms, 422 Analysis
Armstrong’s axiom, 466 CFA, see Confirmatory Factor Analysis
AROMA, see Association Rule CFI, see Comparative Fit Index
Ontology Matching Approach CGF, see cartesian graph of functions
Assess First, 301 CHAID method, 400
Association Rule Ontology Matching characteristic behavioral dimensions,
Approach, 228 300
association rules, 12 chronological conception, 167
ATD, see Anthropological Theory of classes, 26
Didactics classification rules, 401
510 Index

classification trees, 398 educational origin, 353

cognition, 347 EFA, see Exploratory Factor Analysis
cognitive processes, 196 EII, see entropic intensity of implication
Cognitive Tutor, 77 elementary classes, 27
coherence, 369 empirical matrix, 321
cohesion, 24 entropic implication intensity, 20, 22
Comparative Fit Index, 138 entropic intensity of implication, 428,
compartmentalization, 153 452
compartmentalized ways, 147 entropic version, 13
Computer Algebra System, 76 entropy, 428
concept-in-act, 351 episodes, 55
concepts, 235 epistemological origin, 353
conceptual field, 351 Epistemological Representations, 248
conceptual hierarchies, 235 event sequence, 58
Conditional probabilistic reasoning, 175 event types, 58
confidence, 11, 12, 57, 213, 398, 485 evidence illusion, 106
confidence conf (a, b), 16 exceptions, 13
confidence intervals, 168 Expert Bias, 389
Confirmatory Factor Analysis, 132, 136 Exploratory Factor Analysis, 136
conjunction rules, 46
construction process, 196 F-measure, 239
constructivist point of view, 349 Factor analysis, 136
contingency, 14 factorial analysis, 253
contrapositive, 20 Factorial Analysis of Correspondences,
contribution, 33, 45, 325 102
Conversions, 133 fallacy of the transposed conditional,
Correspondence Factor Analysis, 254 167
counter-examples, 11, 14 FARMER, 207
culture, 347 FD, see functional dependency
female form, 125
data analysis method, 11 fictitious individuals, 321
Data Bias, 389 Formalist Axiomatic Geometry, 187
data mining, 11, 12 Freeman-Tukey’s residual, 404
DATALOG, 230 frequency, 60
decision trees, 38 frequency variables, 43
Dep-Miner, 465 frequential, 13
deviance residual, 404 frequential variables, 18
didactic contract, 198 frequentist inference, 164
didactic system, 279 Functional dependencies decomposition,
didactic variables, 249 467
didactical variable, 81 functional dependency, 466
didactics of mathematics, 11, 23 fuzzification, 483
discrete topological C-structure, 32 fuzzy
discrete variables, 43 classes, 483
discursive process, 196 complement, 492
dissimilarity, 29 conclusions, 482
distinct, 143 conjunction, 491
drug discovery, 206 disjunction, 492
dynamic clouds, 43 implication, 493
 Index 511

knowledge discovery, 482 graph, 13, 23

logics, 481 hierarchy, 13, 26
operators, 482 intensity, 14
premises, 482 vector, 31
rules, 482 implicative chains, 121
set, 483 implicative matching relations, 235
variables, 38, 45 inclusion index, 22, 428
independence, 14
Galois lattices, 472 independence hypothesis, 14
gene coregulation, 210 inter-class, 19
Gene expression analysis, 205 inter-class inertia, 33
generativity of the rule, 238 Interactive Learning Environments, 75
generic intensity, 31 interestingness measure, 12, 228
generic pair, 31 interpretation or judgment, 325
genes, 205 interval, 13
genome, 205 interval rank, 208
genotypes, 206 interval variables, 19, 43, 252
geometrical paradigms, 185 IntImp, 427
geometrical working spaces, 185 EII, see entropic intensity of implication
GLUE, 230 REII, see revised EII
GMP-pertinence, 482 TEII, see truncated EII
Goldbach’s conjecture, 255 intra-class, 19
(i)
GPI,theta 433 Intuitionist, 261
graphic conception, 103 IPEE, 432
graphic form, 135 IPEE index, 239
graphic language, 100 itemsets, 422
Graphic representation, 135
Gras’ implication index, 399 J-measure, 56, 385
group typicality, 33 Jaccard measure, 230

H
e ,theta (X)437 KDD, see Knowledge Discovery in
Haberman’s adjusted residual, 404 Databases
HAMB, 207 Knowledge Discovery in Databases, 227,
Hical, 231 463
hierarchical classification, 26
hierarchical clustering, 147 latent variables, 136
hierarchical similarity diagram, 138 leaf, 400
hierarchy tree, 48 Likelihood Linkage Analysis, 14, 239
High ranking, 206 LLA, see Likelihood Linkage Analysis
Historic-epistemological Representa- Loevinger’s coefficient, 16
tions, 248 logic implication, 397
Hypergeometric, 14 Logic of Bayesian inference, 175
logical rule, 14
ILE, see Interactive Learning Environ- low ranking, 206
ments Lukasiewicz’s bounded, 492
IM, see Interestingness Measures
implication intensity, 13, 16 Main Components Analysis, 102
implicative masculine form, 127
distance, 31 material context, 100
512 Index

mathematical modelisation, 325 personality traits, 301

MathXpert, 77 phenotypes, 206
MBTI, see Myers-Briggs Type Indicator Physical Education, 119
meta-rule, 329 Piaget’s model, 349
Method by trials and errors, 269 pitfalls, 384, 449
metric space structure, 32 Poisson process, 61
microarray (DNA chip) technology, 205 Poissonian distribution, 15
microarray analysis, 205 posterior distribution, 167
missing data, 38 pre-experimental analysis, 321
modal, 13 predicted class, 402
modal variables, 18 predictive strategy, 422
Monte Carlo sampling approach, 454 predictor, 399
moon phases, 354 principal component analysis, 192
Morgan’s laws, 492 prior distribution, 167
most typical group, 33 Procedure in natural language, 268
mutual information, 386 profile, 207
Myers-Briggs Type Indicator, 299 propension index, 18
pseudo fuzzy partitions, 483
Natural Axiomatic Geometry, 187 Pseudo-algebraic strategy, 269
Natural Geometry, 187
noise, 12 QL-implications, 494
noise-resistant, 17 quality of a rule, 12
Nominal Variable, 252 quasi-conjunction, 233
non linearly, 12 quasi-equivalence, 233
not symmetrical, 16 quasi-implications, 11, 12
null hypothesis, 451
Numeric Variable, 252 R-implications, 493
numerical variables, 18 R-rule, 13, 24
R-rules of degree 0, 24
objective, 12 rank, 207
objective interestingness measure, 384 recall, 57
objective method, 383 redundancy elimination, 470
obstacles, 352 redundant rule reduction, 38
ontogenetic origin, 352 reference, 350
oPLMap, 230 representations, 119, 131
optimal group, 33 residual terms, 137
optimal typical individual, 31 revised EII, 430
ordinal variables, 43, 252 RMSEA, see Root Mean Square Error
original rules, 47 of Approximation
originality, 46 robust, 12
Osgood’s semantic differentiator, 129 Root Mean Square Error of Approxima-
ostention, 106 tion, 138
OWL ontologies, 228 rule, 11, 12
Rule Bias, 388
p-value, 405, 423, 450 rule discovery, 384
PAPI, 301 rule of rule, 13, 24
PE, see Physical Education Rules Diagnosis, 79
Pearson’s correlation, 16, 210
PerformanSe Echo, 299 S-implications, 493
 Index 513

SAGE, 207 supplementary variables, 45

SBI, see Hical support, 12, 485
Search Bias, 390 Supposed Behaviours, 248
SEM, see Structural equation modeling surprisingness, 14
semantic approaches, 230 symbolic data, 20
semiotic representations, 131 symbolic form, 135
sensitivity, 17 Symbolic representation, 135
sequence repetition, 64
sequential implication intensity, 57, 61 t-conorm, 492
sequential patterns, 55 t-norm, 492
sequential rule, 57, 59 targeting strategy, 422
Shannon’s conditional entropy, 21 temporal sequences, 55
Shannon’s entropy, 20 terminological approaches, 230
SIA, see Statistical Implicative Analysis terms, 235
signification, 350 test of χ2 , 365
significative levels, 13 test value, 453
signifiers, 350 test-value percent principle, 450
SII, see sequential implication intensity theorem-in-act, 351
similarity, 14 theoretical model, 14
analysis, 107 threshold, 14
index, 49 transitive closures, 50
intensity, 42 tree, 401
tree, 42, 48 truncated EII, 430
skills reconstruction, 285 tumour classification, 206
socio-cognitive conflict, 359 TVpercent, 453
socio-cultural obstacles, 361 typicality, 32, 45, 50, 325
software CHIC, 23
Sosie, 301 ultrametric inequality, 28
SPAD_T, 364 unlikelihood, 14
SPSS, 255 unsupervised analysis, 205
standardized residual, 403
statistical implication, 397 Variables on intervals, 19
Statistical Implicative Analysis, 11 variables over intervals, 43
statistical interestingness measures, 421 vectorial data, 38
structural approaches, 230 verbal form, 135
Structural equation modeling, 136 verbal representation, 135
student model, 75 visualization process, 196
subjective, 12
subjective method, 383 window, 59
supervised analysis, 205
supplementary individuals, 13 Zadeh’s minimum, 492