Applications of Machine Learning and Data Analytics

Models in Maritime Transportation

Applications of Machine Learning

Machine learning and data analytics can be used to inform technical, commercial and financial

Transportation
Analytics Models in Maritime
Applications of Machine Learning and Data
decisions in the maritime industry. Applications of Machine Learning and Data Analytics
Models in Maritime Transportation explores the fundamental principles of analysing maritime-
transportation related practical problems using data-driven models, with a particular focus on
machine learning and operations research models.
and Data Analytics Models in
Data-enabled methodologies, technologies, and applications in maritime transportation are
clearly and concisely explained, and case studies of typical maritime challenges and solutions
Maritime Transportation
are also included. The authors begin with an introduction to maritime transportation, followed
by chapters providing an overview of ship inspection by port state control, and the principles
of data driven models. Further chapters cover linear regression models, Bayesian networks,
support vector machines, artificial neural networks, tree-based models, association rule
learning, cluster analysis, classic and emerging approaches to solving practical problems in
maritime transport, incorporating shipping domain knowledge into data-driven models,
explanation of black-box machine learning models in maritime transport, linear optimization,
advanced linear optimization, and integer optimization. A concluding chapter provides an
overview of coverage and explores future possibilities in the field.
Ran Yan and Shuaian Wang
The book will be especially useful to researchers and professionals with expertise in maritime
research who wish to learn how to apply data analytics and machine learning to their fields.

About the Authors

Ran Yan is a research assistant professor in the Department of Logistics and Maritime Studies
at The Hong Kong Polytechnic University, China.

Shuaian Wang is a professor in the Department of Logistics and Maritime Studies at The
Hong Kong Polytechnic University, China.

Yan and Wang

The Institution of Engineering and Technology

theiet.org
978-1-83953-559-8
IET TRANSPORTATION SERIES 38

Applications of Machine
Learning and Data Analytics
Models in Maritime
Transportation
Other related titles:

Volume 1 Clean Mobility and Intelligent Transport Systems M. Fiorini and J-C. Lin (Editors)
Volume 2 Energy Systems for Electric and Hybrid Vehicles K.T. Chau (Editor)
Volume 5 Sliding Mode Control of Vehicle Dynamics A. Ferrara (Editor)
Volume 6 Low Carbon Mobility for Future Cities: Principles and Applications H. Dia (Editor)
Volume 7 Evaluation of Intelligent Road Transportation Systems: Methods and Results M. Lu (Editor)
Volume 8 Road Pricing: Technologies, economics and acceptability J. Walker (Editor)
Volume 9 Autonomous Decentralized Systems and their Applications in Transport and Infrastructure
K. Mori (Editor)
Volume 11 Navigation and Control of Autonomous Marine Vehicles S. Sharma and B. Subudhi (Editors)
Volume 12 EMC and Functional Safety of Automotive Electronics K. Borgeest
Volume 15 Cybersecurity in Transport Systems M. Hawley
Volume 16 ICT for Electric Vehicle Integration with the Smart Grid N. Kishor and J. Fraile-Ardanuy (Editors)
Volume 17 Smart Sensing for Traffic Monitoring Nobuyuki Ozaki (Editor)
Volume 18 Collection and Delivery of Traffic and Travel Information P. Burton and A. Stevens (Editors)
Volume 20 Shared Mobility and Automated Vehicles: Responding to socio-technical changes and pandemics Ata
Khan and Susan Shaheen
Volume 23 Behavioural Modelling and Simulation of Bicycle Traffic L. Huang
Volume 24 Driving Simulators for the Evaluation of Human-Machine Interfaces in Assisted and Automated
­Vehicles T. Ito and T. Hirose (Editors)
Volume 25 Cooperative Intelligent Transport Systems: Towards high-level automated driving M. Lu (Editor)
Volume 26 Traffic Information and Control Ruimin Li and Zhengbing He (Editors)
Volume 30 ICT Solutions and Digitalisation in Ports and Shipping M. Fiorini and N. Gupta
Volume 32 Cable Based and Wireless Charging Systems for Electric Vehicles: Technology and control,
management and grid integration R. Singh, S. Padmanaban, S.Dwivedi, M. Molinas and F. Blaabjerg (Editors)
Volume 34 ITS for Freight Logistics H. Kawashima (Editor)
Volume 36 Vehicular ad hoc Networks and Emerging Technologies for Road Vehicle Automation A. K. Tyagi and
S Malik
Volume 38 The Electric Car M.H. Westbrook
Volume 45 Propulsion Systems for Hybrid Vehicles J. Miller
Volume 79 Vehicle-to-Grid: Linking Electric Vehicles to the Smart Grid J. Lu and J. Hossain (Editors)
Applications of Machine
Learning and Data Analytics
Models in Maritime
Transportation
Ran Yan and Shuaian Wang

The Institution of Engineering and Technology

Published by The Institution of Engineering and Technology, London, United Kingdom
The Institution of Engineering and Technology is registered as a Charity in England &
Wales (no. 211014) and Scotland (no. SC038698).
© The Institution of Engineering and Technology 2022
First published 2022

This publication is copyright under the Berne Convention and the Universal Copyright
Convention. All rights reserved. Apart from any fair dealing for the purposes of research or
private study, or criticism or review, as permitted under the Copyright, Designs and Patents
Act 1988, this publication may be reproduced, stored or transmitted, in any form or by
any means, only with the prior permission in writing of the publishers, or in the case of
reprographic reproduction in accordance with the terms of licences issued by the Copyright
Licensing Agency. Enquiries concerning reproduction outside those terms should be sent to
the publisher at the undermentioned address:
The Institution of Engineering and Technology
Futures Place
Kings Way, Stevenage
Herts, SG1 2UA, United Kingdom
www.theiet.org
While the authors and publisher believe that the information and guidance given in this
work are correct, all parties must rely upon their own skill and judgement when making use
of them. Neither the author nor publisher assumes any liability to anyone for any loss or
damage caused by any error or omission in the work, whether such an error or omission is
the result of negligence or any other cause. Any and all such liability is disclaimed.
The moral rights of the author to be identified as author of this work have been asserted by
him in accordance with the Copyright, Designs and Patents Act 1988.

British Library Cataloguing in Publication Data

A catalogue record for this product is available from the British Library

ISBN 978-1-83953-559-8 (hardback)

ISBN 978-1-83953-560-4 (PDF)

Typeset in India by Exeter Premedia Services Private Limited

Printed in the UK by CPI Group (UK) Ltd, Croydon
Cover Image: Longhua Liao/ Moment via Getty Images
 Contents

About the Authors ix

1 Introduction of maritime transportation 1
1.1 Overview of maritime transport 1
1.2 World fleet structure 1
1.2.1 Bulk carrier 1
1.2.2 Oil tanker 2
1.2.3 Container ship 2
1.3 Key roles in the shipping industry 3
1.3.1 Ship owner 3
1.3.2 Ship operator 3
1.3.3 Ship management company 3
1.3.4 Flag state 3
1.3.5 Classification society 4
1.3.6 Charterer 4
1.3.7 Freight forwarder 4
1.3.8 Ship broker 5
1.4 Container liner shipping 5

2 Ship inspection by port state control 9

2.1 Key issues in maritime transport 9
2.1.1 Maritime safety management 9
2.1.2 Marine pollution control 10
2.1.3 Seafarers’ management 10
2.2 Port state control 10
2.2.1 The background and development of PSC 11
2.2.2 Ship selection in PSC 11
2.2.3 Onboard inspection procedure 12
2.2.4 Inspection results 14
2.3 Data set used in this book 15

3 Introduction to data-­driven models 21
3.1 Predictive problem and its application in maritime transport 21
3.1.1 Introduction of predictive problem 21
3.1.2 Examples of predictive problem in maritime transport 22
vi  Machine learning and data analytics for maritime studies

3.1.3 Comparison of theory-based modeling and data-driven

modeling23
3.1.4 Popular data-driven models 23

4 Key elements of data-­driven models 29

4.1 Comparison of three popular data-driven models 29
4.2 Procedure of developing ML models to address maritime transport
problems29
4.2.1 Problem specification 32
4.2.2 Feasibility assessment 32
4.2.3 Data collection 32
4.2.4 Feature engineering 34
4.2.5 Model construction 43
4.2.6 Model refinement 48
4.2.7 Model assessment, interpretation/explanation, and conclusion 50

5 Linear regression models 51
5.1 Simple linear regression and the least squares 51
5.2 Multiple linear regression 53
5.3 Extensions of multiple linear regression 55
5.3.1 Polynomial regression 55
5.3.2 Logistic regression 56
5.4 Shrinkage linear regression models 59
5.4.1 Ridge regression 60
5.4.2 LASSO regression 61

6 Bayesian networks 63
6.1 Naive Bayes classifier 63
6.2 Semi-naive Bayes classifiers 68
6.3 BN classifiers 73

7 Support vector machine 79
7.1 Hard margin SVM 79
7.2 Soft margin SVM 83
7.3 Kernel trick 86
7.4 Support vector regression 90

8 Artificial neural network 93
8.1 The structure and basic concepts of an ANN 93
8.1.1 Training of an ANN model 97
8.1.2 Hyperparameters in an ANN model 100
8.2 Brief introduction of deep learning models 103
 Contents  vii

9 Tree-­based models 105

9.1 Basic concepts of a decision tree 105
9.2 Node splitting in classification trees 106
9.2.1 Iterative dichotomizer 3 (ID3) 106
9.2.2 C4.5 109
9.2.3 Classification and regression tree (CART) 110
9.2.4 Node splitting in regression trees 113
9.3 Ensemble learning on tree-based models 115
9.3.1 Bagging 117
9.3.2 Boosting 120

10 Association rule learning 129

10.1 Large item sets 129
10.2 Apriori algorithm 131
10.3 FP-growth algorithm 139

11 Cluster analysis 143

11.1 Distance measure in clustering 143
11.1.1 Distance measure of examples 143
11.1.2 Distance measure of clusters 146
11.2 Metrics for clustering algorithm performance evaluation 147
11.2.1 Clustering algorithms 149
11.2.2 K-means (partition-based method) 149
11.2.3 DBSCAN (density-based method) 151
11.2.4 Agglomerative algorithm (hierarchy-based methods) 154

12 Classic and emerging approaches to solving practical problems

in maritime transport 161
12.1 Topics in maritime transport research 161
12.2 Research methods and their specific applications to maritime
transport research 163
12.3 Issues of adopting data-driven models to address problems
in maritime transportation 167
12.3.1 Data 168
12.3.2 Model 169
12.3.3 User 170
12.3.4 Target 172

13 Incorporating shipping domain knowledge into data-­driven models 179

13.1 Considering feature monotonicity in ship risk prediction 180
13.1.1 Introduction of monotonicity in the ship risk prediction
problem180
13.1.2 Integration of monotonic constraint into XGBoost 182
viii  Machine learning and data analytics for maritime studies

13.2 Integration of convex and monotonic constraints into ANN

(artifical neural network) 184

14 Explanation of black-­box ML models in maritime transport 191

14.1 Necessity of black-box ML model explanation in the maritime
industry191
14.1.1 What is the explanation for ML models 191
14.1.2 Propose and evaluate explanations for black-box ML models 194
14.2 Popular methods for black-box ML model explanation 196
14.2.1 Forms and types of explanations 196
14.2.2 Introduction of intrinsic explanation model using DT
as an example198
14.2.3 SHAP method 200

15 Linear optimization 209

15.1 Basics 209
15.2 Classification of linear optimization models according to solutions 213
15.3 Equivalence between different formulations 215
15.4 Graphical method for models with two variables 218
15.5 Using software to solve linear optimization models 225
15.6 An in-depth understanding of linear optimization 227
15.7 Useful applications of linear optimization solvers 228

16 Advanced linear optimization 231

16.1 Network flow optimization 231
16.2 Dummy nodes and links 245
16.3 Using linear formulations for nonlinear problems 248
16.4 Practice 250

17 Integer optimization 275

17.1 Formulation I: natural integer decision variables 275
17.2 Formulation II: 0–1 decision variables 278
17.3 Formulation III: complex logical constraints 281
17.4 Solving mixed-integer optimization models 282
17.5 Formulation IV: challenging problems 283
17.6 Formulation V: linearizing binary variables multiplied by another
variable288
17.7 Practice 289

18 Conclusion297
18.1 Summary of this book 297
18.2 Future research agenda 298

Index301
 About the Authors

Ran Yan is a research assistant professor in the Department of Logistics and

Maritime Studies at The Hong Kong Polytechnic University (PolyU), China. Dr. Yan
received her Bachelor of Science degree from Hohai University in China in 2018 and
her Master of Philosophy and Doctor of Philosophy degrees from The Hong Kong
Polytechnic University in 2020 and 2022, respectively. Dr. Yan’s research interests
include applying data analytics methods and technologies to improve shipping effi-
ciency and green shipping management. Dr. Yan has published more than 30 papers in
international journals and conference proceedings, such as Transportation Research
Part B/C/E, Transport Policy, Journal of Computational Science, Maritime Policy &
Management, Ocean Engineering, Engineering, Sustainability, and Electronic
Research Archive, and won several times of best paper/student paper award from
international conferences. Dr. Yan is an editorial assistant of Cleaner Logistics and
Supply Chain.

Shuaian Wang is currently Professor at The Hong Kong Polytechnic University

(PolyU), China. Prior to joining PolyU, he worked as a faculty member at Old
Dominion University, USA, and the University of Wollongong, Australia. Dr. Wang’s
research interests include big data in shipping, green shipping, shipping operations
management, port planning and operations, urban transport network modeling, and
logistics and supply chain management. Dr. Wang has published over 200 papers
in journals such as Transportation Research Part B, Transportation Science, and
Operations Research. Dr. Wang is an editor-in-chief of Cleaner Logistics and
Supply Chain and Communications in Transportation Research, an associate editor
of Transportation Research Part E, Flexible Services and Manufacturing Journal,
Transportmetrica A, and Transportation Letters, a handle editor of Transportation
Research Record, an editorial board editor of Transportation Research Part B, and
an editorial board member of Maritime Transport Research. Dr. Wang dedicates to
rethinking and proposing innovative solutions to improve the efficiency of maritime
and urban transportation systems, to promote environmental friendly and sustainable
practices, and to transform business and engineering education.
This page intentionally left blank
 TO MY PARENTS
IN MEMORY OF MY BELOVED MATERNAL GRANDPARENTS
—Ran

TO MY DAUGHTER AMY
—Shuaian
This page intentionally left blank
 Chapter 1
Introduction of maritime transportation

1.1 Overview of maritime transport

Maritime transportation is the transport of passengers and cargoes by sea which can
be dated back to ancient Egypt times. It is the backbone of global trade, manufac-
turing supply chain, and world economy, as about 80% of world trade by volume
is carried out by ocean-­going vessels [1]. Even during the ever-­hard time brought
about by the COVID-­19 pandemic, the maritime transport still plays a key role in
moving various goods and products, especially the necessary living and medical
supplies, around the world. The world economic development has a critical impact
on maritime transport. According to the review of maritime transport produced by
the United Nations Conference on Trade and Development (UNCTAD), the interna-
tional maritime trade volumes from 1970 to 2020 are shown in Table 1.1 [1].

1.2  World fleet structure

The total number of ships of 100 gross tons and above reached 99 800 in 2020.
Different types of ships are responsible for carrying different types of commodities.
The top three ship types with the largest market shares in the world fleet by dead-
weight tonnage (DWT) are bulk carriers (913 032 000 tons, 42.77%), oil tankers
(619 148 000 tons, 29.00%), and container ships (281 784 000 tons, 13.20%) in
2021, according to the UNCTAD based on data from Clarksons Research [1]. A brief
introduction of the three types of ships is given in the following subsections.

1.2.1 Bulk carrier
Bulk carriers mainly transport dry cargoes in bulk quantities in a loose manner,
that is, the cargoes are free from any specific packaging. Major category of bulk
carriers as per size includes handysize carriers (usually between 25 000 and 40 000
DWT), handymax carriers (usually between 40 000 and 60 000 DWT), Panamax
carriers (usually between 60 000 and 100 000 DWT), post-­Panamax carrier (usually
between 80 000 and 120 000 DWT), capesize carrier (usually between 100 000 and
200 000 DWT), and very large bulk carrier (usually over 200 000 DWT). Bulk car-
riers are mainly used to transport ores, coals, cement, etc.
2  Machine learning and data analytics for maritime studies

Table 1.1  International maritime trade from 1970 to 2020 (millions of tons
loaded)

Year Tanker trade* Main bulk† Other dry Total (all

cargo‡ cargoes)
1970 1 440 448 717 2 605
1980 1 871 608 1 225 3 704
1990 1 755 988 1 265 4 008
2000 2 163 1 186 2 635 5 984
2005 2 422 1 579 3 108 7 109
2010 2 752 2 232 3 423 8 408
2011 2 785 2 364 3 626 8 775
2012 2 840 2 564 3 791 9 195
2013 2 828 2 734 3 951 9 513
2014 2 825 2 964 4 054 9 842
2015 2 932 2 930 4 161 10 023
2016 3 058 3 009 4 228 10 295
2017 3 146 3 151 4 419 10 716
2018 3 201 3 215 4 603 11 019
2019 3 163 3 218 4 690 11 071
2020 2 918 3 181 4 549 10 648
Source: UNCTAD secretariat, based on reports from countries, governments, port industries websites,
and specialist sources.
*Includes crude oil, refined petroleum products, gas, and chemicals.
†
Includes iron ore, grain, and coal.
‡
Includes minor bulk commodities, containerized trade, and residual general cargo.

1.2.2 Oil tanker
Oil tanker is a type of tanker ship which is designed for the particular purpose
of transporting liquefied goods, mainly including crude oil and its by-­products.
According to their sizes, oil tankers can be divided into Panamax (usually
between 50 000 and 75 000 DWT), Aframax (coming from the Average Freight
Rate Assessment system, usually between 75 000 and 120 000 DWT), Suezmax
(usually between 120 000 and 180 000 DWT), very large crude carrier (usually
between 200 000 and 320 000 DWT), and ultra large crude carrier (usually more
than 320 000 DWT).

1.2.3 Container ship
Container ships have specially designed structures to hold a large quantity of
cargoes compacted in different types of containers. Container ship capacity is
measured in twenty-­foot equivalent units (TEUs). They can be further divided
into feeder (1 001–2 000 TEU), feedermax (2 001–3 000 TEU), Panamax (3
001–5 100 TEU), post-­Panamax (5 101–10 000 TEU), new Panamax (10 000–
14 500 TEU), and ultra large container vessel (over 14 500 TEU). Container
 Introduction of maritime transportation  3

shipping is revolutionary in the form of cargoes that are ferried and transported
around the world, as it is the foundation of international multimodal transport
that connects ship with truck and rail. Globalization would not have been pos-
sible without containerization. Common cargoes transported by container ships
include manufactured goods, raw materials, food goods, agricultural products,
and seafood in refrigerated containers. More information on container liner
shipping is provided in section 1.4.

1.3 Key roles in the shipping industry

1.3.1 Ship owner
The owner of a merchant ship is called a ship owner, which can be companies, people, and
investment funds. The ownership is achieved by building new ships or purchasing second-­
hand ships. Technical or commercial operations are provided by some ship owners, while
others prefer to outsource these services to specialized companies. The party that offers
commercial management service is called a ship operator, and the party that offers techni-
cal management service is called a ship management company.

1.3.2 Ship operator
Ship operator is the company in charge of daily matters on ships. It needs to tackle
with several tactical problems such as determining the cargo orders as well as the
freight rate, fleet deployment, ship routing and scheduling, and port service negotia-
tion. In addition, ship operator also needs to deal with operational problems such as
deciding ship loading, selecting sailing speeds, routes, and bunkering ports.

1.3.3 Ship management company

Ship owners can choose to delegate various managerial functions to a third
party, which is the so-­called ship management company or simply the manager.
The delegated services of a management company include technical manage-
ment, commercial management, and crew management. To be more specific,
technical management services include but are not limited to insuring a ship
that complies with international codes and conventions, arranging repairs and
maintenance, and providing technical consultants. Crew management services
include crew recruitment, training, insurance, payroll, etc. Ship managers also
offer commercial management services, such as securing vessel employment,
obtaining marine insurance, and providing risk management services. One ship
management company can manage a large number of fleets on behalf of numer-
ous ship owners.

1.3.4 Flag state
The ship flag state is the nationality of a ship under whose laws the ship is ply-
ing in the open sea. A flag state is responsible for conducting regular inspections
of its ships to ensure the safety of its cargo and crew (which is called flag state
4  Machine learning and data analytics for maritime studies

control), collect taxes, and regulate the pollution levels by imposing maritime
policies and laws. In turn, ships receive protections and preferential treatments
such as tax, certification, trade rights, and security from their flags of registra-
tion. The flag state of a ship plays an important role in many aspects of ship
management and operation, such as vessel leasing, sale and purchase, newbuild-
ing deliveries, financing, and different priorities of owners and mortgagees. The
practice of registering a ship to a state different from that of the ship owner is
known as the flag of convenience.

1.3.5  Classification society

Classification societies are mainly responsible for setting technical standards
for seagoing cargo vessels based on their experience, expertise, and research
in the shipping industry while considering international maritime conventions
and regulations. A classification society should also inspect the ships under its
classing ever since the ship construction process. The periodic operation of their
ships should also be provided to ensure that they continuously meet the related
regulations. Classification society plays an important role in enhancing maritime
safety and protecting the marine environment, and their work interacts with ship
flag states, port state control, surveyors, and among other groups in addition to
ship owners.

1.3.6 Charterer
Ship charterer is a company that hires ships from ship owners to transport their
cargoes, and the contract between them is known as a charterparty. Time charter
and voyage charter are the most popular charter party agreements, where the
ship owner rents a ship to the charterer for a jointly determined period with
restrictions on trading and cargoes imposed in the former, while a ship is hired to
carry a particular cargo between specific places in the latter. Usually, under the
time charterparty, the ship owner takes care of their crew, maintains ship stores
and provisions, and conducts ship maintenance. Meanwhile, the time charterers
are responsible for fuel arrangement, port and canal dues, wharfage, and many
other types of operational costs. Whereas in a typical voyage charterparty, most
of the operational costs are covered by the ship owner, and the voyage charterer
is responsible for the costs in ports and terminals of the specific voyage. Another
popular charterparty is a bareboat charter, where a ship charterer hires a ship
without administration or technical maintenance provided, and it bears the full
operating expenses. In all cases, both ship owners and charterers are free to
negotiate their liabilities.

1.3.7 Freight forwarder
A freight forwarder is a person or an entity that manages shipments to transport the goods
from origin to destination, and it takes several duties which could include consultancy on
export and/or import costs and regulations, arrangement of goods transportation, cargo
insurance, document translation, and communication with clients.
 Introduction of maritime transportation  5

1.3.8 Ship broker
Ships are chartered every day through ship brokers to carry cargoes worldwide. A
ship owner’s broker takes the responsibility to arrange the most profitable employ-
ment of the vessels under the owner’s control, and the ship charterer’s broker helps
the charter to find the most effective way to transport the cargoes with the right ship
at the right price. Therefore, it is necessary that the ship broker to be professional
and is aware of the current freight market and is also able to forecast its future status
as well as acts as the bridge between the ship owner and the charterer.

1.4 Container liner shipping

A majority of cargoes in supermarkets, such as fruits and vegetables, kitchen appli-

ances, furniture, garments, meats, fish, dairy products, and toys, are transported in
containers by ship. Containers are usually expressed in terms of TEUs, a box that
is 20 feet long (6.1 m). Throughout this book, unless otherwise specified, we use
“TEU” to express “the number of containers” or “the volume of containers.”
Containers are transported by ship on liner services, which are similar to bus ser-
vices. Figure 1.1 is the Central China 2 (CC2) service operated by Orient Overseas
Container Line (OOCL), a Hong Kong-­based shipping company. We call it a ser-
vice, a route, or a service route. A route is a loop, and the port rotation of a route is
the sequence of ports of call on the route. Any port of call can be defined as the first
port of call. For example, if we define Ningbo as the first port of call, then Shanghai
is the second port of call, and Los Angeles is the third port of call. We can therefore
represent the port rotation of the route as follows:

‍ Ningbo (1) ! Shanghai (2) ! Los Angeles (3) ! Ningbo (1)‍

Figure 1.1   CC2 service by OOCL

6  Machine learning and data analytics for maritime studies

Figure 1.2   CC1 service by OOCL

Note that on a route, different ports of call may be the same physical port. For
example, the Central China 1 (CC1) service of OOCL shown in Figure 1.2 has the
port rotation below:
Shanghai (1) ! Kwangyang (2) ! Pusan (3) ! Los Angeles (4) !
‍ Oakland (5) ! Pusan (6) ! Kwangyang (7) ! Shanghai (1) ‍
Both the second and the seventh ports of call are Kwangyang, and both the third and
the sixth ports of call are Pusan.
A leg is the voyage from one port of call to the next. Leg i‍ ‍is the voyage from the i‍
‍th port of call to port of call i‍ + 1‍. The last leg is the voyage from the last port of call to
the first port of call. On CC1, the second leg is the voyage from Kwangyang (the second)
to Pusan (the third), and the seventh leg is the voyage from Kwangyang (the seventh) to
Shanghai (the first).
The rotation time of a route is the time required for a ship to start from the
first port of call, visit all ports of call on the route, and return to the first port
of call. As can be read from Figures 1.1 and 1.2, the rotation time of CC2 is 35
days*, and the rotation time of CC1 is 42 days. Each route provides a weekly fre-
quency, which means that each port of call is visited on the same day every week.
Therefore, a string of five ships are deployed on CC2, and the headway between
two adjacent ships is 7 days. These five ships usually have the same TEU capacity

13 days from the departure from Shanghai to the arrival at Los Angeles + 17 days from the departure
* 

from Los Angeles to the arrival at Shanghai + 2 days spent at the port of Shanghai (Wednesday to
Friday) + 3 days spent at the port of Los Angeles (Thursday to Sunday).
 Introduction of maritime transportation  7

and other characteristics. Unless otherwise specified, we assume weekly frequen-

cies for all routes.
On CC2, it may be the case that we are interested in container transportation
from Ningbo to Los Angeles, Shanghai to Los Angeles, Los Angeles to Ningbo, and
Los Angeles to Shanghai, and there is no container to transport between Ningbo and
Shanghai. We therefore define an origin-­destination (OD) pair as an ordered pair of
origin port and destination port. In this example, there are four OD pairs: (Ningbo
and Los Angeles), (Shanghai and Los Angeles), (Los Angeles and Ningbo), and
(Los Angeles and Shanghai). The demand for an OD pair is the number of TEUs
per week to transport. For example, if the demand for the OD pair (Ningbo and Los
Angeles) is 5 000, then there are 5 000 TEUs from potential customers to be trans-
ported from Ningbo to Los Angeles every week. Nevertheless, the shipping line may
choose to transport, e.g., 4 675 TEUs and reject the remaining 325 TEUs.
Unless otherwise specified, throughout this book we examine tactical planning
problems rather than operational problems. For example, we might predict that in
the next 6 months, there will be 2 000 TEUs per week to transport from Ningbo
to Los Angeles, and we aim to estimate the profit we can make if the CC2 service
is operated. As the “2 000 TEUs per week to transport” may not be accurate, it is
generally not a problem to say “1845.3 TEUs will be transported”(a more detailed
analysis can be found in Reference 2). In other words, we assume that the number of
TEUs is a real number rather than an integer, unless otherwise specified.

References

[1] ‘United Nations conference on trade and development’ in Review of Maritime

Transport. New York, NY: United Nations Publications; 2022.
[2] Wang S. ‘Essential elements in tactical planning models for container liner
shipping’. Transportation Research Part B. 2013;54:84–99.
This page intentionally left blank
 Chapter 2
Ship inspection by port state control

2.1 Key issues in maritime transport

2.1.1 Maritime safety management

Maritime transport is a highly globalized industry in terms of operation and manage-
ment. For ship operation, ocean-­going vessels sail on the high seas from the origin
port in one country/region to the destination port in another country/region. For ship
management, parties responsible for ship ownership, crewing, and operating may
locate in different countries and regions. Even the country of registration, i.e., ship
flag state, may not have a direct link and connection with a ship’s activities as the
ship may not frequently visit the ports belonging to its flag state. For inland coun-
tries such as Mongolia, the ships registered under it never visit its ports. Such com-
plex and disintegrated nature of the shipping industry makes it hard to control and
regulate international shipping activities, and thus pose danger to maritime safety,
the marine environment, and the crew and cargoes carried by ocean-­going vessels.
Shipping is one of the world’s most dangerous industries due to the complex and
ever-­changing environment at sea, the dangerous goods carried, and the difficulties in
search and rescue. Safety at sea is always put at the highest priority in ship operation and
management. It is widely believed that the most effective and efficient way of improv-
ing safety at sea is to develop international regulations that should be followed by all
shipping nations [1]. A unified and permanent international body was expected to be
established for regulation and supervision by several nations from the mid-­19th century
onward, and the hopes came true after the International Maritime Organization (IMO,
whose original name was Inter-­Governmental Maritime Consultative Organization)
was established at an international conference in Geneva held in 1948. Through hard
efforts of all parties, the members of IMO met for the first time in 1959, one year after
the IMO convention came into force. The IMO’s task was to adopt a new version of
the most important conventions on maritime safety, i.e., the International Convention
for the Safety of Life at Sea, which specifies minimum safety standards for ship con-
struction, equipment, and operation. It covers comprehensive aspects of shipping safety,
including vessel construction, fire safety, life-­saving arrangements, radio communica-
tions, navigation safety, cargo carriage, dangerous goods transporting, the mandatory of
the International Safety Management (ISM) code, verification of compliance, and mea-
sures for specific ships, and is constantly amended [2]. The Maritime Safety Committee
10  Machine learning and data analytics for maritime studies

is responsible for every aspect of maritime safety and security, and it is the highest
technical body of the IMO.

2.1.2 Marine pollution control

Although safety has always been the first priority and the most important responsi-
bility, pollution caused by maritime accidents and shipping activities began to merge
and is receiving an increasing attention from the parties. The most important inter-
national convention is the International Convention for the Prevention of Pollution
from Ships proposed in 1973 and modified in 1978 (MARPOL 73/78), which aims
at preventing pollution of the marine environment caused by vessel operations and
accidents. As of January 2022, six technical annexes are included in the MARPOL
convention, including prevention of pollution by oil, noxious liquid substances in
bulk, harmful substances in packaged form, sewage from ships, garbage from ships,
and air pollution from ships [3]. The Marine Environment Protection Committee of
the IMO takes the full responsibility to address matters concerned with prevention
and control of pollution from ships by adopting and amending the related conven-
tions and regulations and ensure their enforcement.

2.1.3 Seafarers’ management
Seafarers are regarded as key workers by the IMO, who are essential to shipping and
the world, in particular, during the COVID-­19 pandemic. They are at the core of ship-
ping’s future. Basic requirements on the training, certification, and watchkeeping of
seafarers on an international level are established in the International Convention on
Standards of Training, Certification, and Watchkeeping for Seafarers by the IMO,
which entered into force in 1984 [4]. Besides, the milestone to ensure a decent liv-
ing and working environment for seafarers is the Maritime Labor Convention, 2006
(MLC2006), which entered into force in 2013 and is developed by the International
Labor Organization [5]. Minimum requirements for seafarers regarding conditions
of employment, work, accommodation, food and catering, health care and medi-
cal care onboard, as well as welfare and social security protection are specified in
MLC2006. A sub-­committee on Human Element, Training, and Watchkeeping deals
with the human sides of shipping in the IMO.

2.2 Port state control

A ship whose hull, machinery, equipment, or operational safety is substantially below the
standards required by the relevant convention or whose crew is not in conformity with the
safe manning document is deemed as a substandard ship [6]. The ship flag state is the first
line of defense against substandard shipping. It is obliged to ensure its ships are periodi-
cally surveyed and re-­certified by carrying out inspections by its own surveyors or by an
authorized party called a recognized organization (RO). However, due to the international-
ism of shipping activities, many ships do not regularly call at ports in their flag states, and
thus restrict the role that flag state plays in identifying substandard ships.
 Ship inspection by port state control  11

2.2.1 The background and development of PSC

PSC is then developed, aiming to inspect foreign visiting ships in national ports to verify
that “the condition of the ship and its equipment comply with the requirements of inter-
national regulations and that the ship is manned and operated in compliance with these
rules” as mentioned by the IMO [7]. During an inspection, a condition onboard that
does not comply with the requirements of the relevant convention is called a deficiency.
The number and nature of the deficiencies found onboard determine the corresponding
action taken by the PSC officer(s) (PSCO[s]). Common actions include rectifying a defi-
ciency at the next port within 14 days or before departure and ship detention. Especially,
ship detention is an intervention action taken by the port state that prevents a severely
substandard ship from proceeding to sea until it would not present danger to the ship or
persons onboard as well as to the marine environment.
PSC inspection is carried out on a regional level. The Memorandum of Understanding
(MoU) on PSC was first signed in 1982 by 14 European countries, which is called the
Paris MoU and marks the establishment of PSC. Since then, the number of member
states of the Paris MoU has constantly increased, and it contains 27 participating mari-
time administrations covering the waters of the European coastal States and the North
Atlantic basin as of January 2022. Another large regional MoU is in the Far East respon-
sible for the Asia Pacific region, which is called Tokyo MoU and was signed in 1993.
It now contains 22 member states. In addition, there are another seven MoUs on PSC,
namely Acuerdo de Viña del Mar (Latin America), Caribbean MoU (Caribbean), Abuja
MoU (West and Central Africa), Black Sea MoU (the Black Sea region), Mediterranean
MoU (the Mediterranean), Indian Ocean MoU (the Indian Ocean), and the Riyadh MoU.
The main objectives of constructing MoUs are constructing an improved and harmo-
nized PSC system, strengthening cooperation and information exchange among mem-
ber states, and avoiding multiple inspections within a short period. Apart from the nine
regional MoUs, the United States Coast Guard maintains the tenth PSC regime.

2.2.2 Ship selection in PSC

One critical step in ship inspection by PSC is the identification and selection of ships
with a higher risk of inspection. This is because on the one hand, not all foreign visit-
ing ships are able to be inspected due to the limited inspection resources, while on
the other hand, not all ships are substandard. A PSC inspection is usually conducted
by one or two inspectors, i.e., PSCOs, and are authorized by the port state to carry
out PSC inspections. A typical PSC inspection usually lasts 2–3 hours. According to
the estimation of Knapp [8], the average cost of an inspection free from deficiency
is US$506 for the port. If any deficiency is found, the average cost per inspec-
tion increases to US$759. Therefore, carrying out an inspection can be both time-­
consuming and costly. Meanwhile, among all the foreign visiting ships, only part of
them are substandard, or in other words, with a deficiency or deficiencies detected.
For example, in the Tokyo MoU, among all the inspections between 2009 and 2019,
about 40% of them are without deficiency detected, while the average detention rate
is 4.18% [9]. It needs to be mentioned that both statistics are calculated based only
on the ships selected for inspection. If all the foreign visiting ships are considered,
12  Machine learning and data analytics for maritime studies

both rates will be lower (as only ships with relatively high risk are selected for
inspection). Therefore, accurately identifying substandard ships with a higher risk,
i.e., with a larger number of deficiencies or a higher probability of detention, is of
vital importance to enhance the effectiveness of PSC.
To achieve efficient PSC inspection, uniform ship selection methods are adopted
by individual MoUs, while the methods adopted by different MoUs might be dif-
ferent. One fundamental ship selection method is based on expert experience. For
example, in the Mediterranean MoU, ships can be exempted from further inspection
if they have been inspected within the last 6 months and found to comply with the
regulations [10]. A more advanced ship selection method is called the new inspec-
tion regime (NIR), which is adopted by the Abuja MoU, the Black Sea MoU, the
Paris MoU, and the Tokyo MoU. The NIR takes ship age and type, performance of
ship flag/RO/company*, and the number of deficiencies and detentions in the last 3
years to calculate ship risk profile (SRP). An example of the information sheet for
ship risk calculation used in the Tokyo MoU is presented in Table 2.1. Under the
NIR, ships are divided into three risk profiles in accordance with their risk levels
calculated, namely high-­risk ships (HRS), standard-­risk ships (SRS), and low-­risk
ships (LRS). Inspection time window is then attached to each SRP as shown in
Table 2.2. It is also noted that the NIR for ship risk calculation can be more or less
different in different MoUs, including the information sheet used and the time win-
dows attached.

2.2.3 Onboard inspection procedure

According to the guidance of the IMO, a general PSC inspection can include an initial
inspection and a more detailed inspection. An initial inspection may start before a PSCO
gets on board a ship by observing its appearance to gain an impression of the ship’s
overall maintenance standards, like the paintwork, corrosion or pitting, and unrepaired
damage. After getting onboard, the PSCO checks and examines the ship’s relevant docu-
ments and certificates in accordance with the ship’s type, size, and year of build, as well
as the certificates and documents of its crew. Then, the PSCO walks around the ship in an
unsupervised manner to inspect its overall condition, including its equipment, navigation
bridge, forecastle, cargo holds/areas, engine room, and pilot transfer arrangements. If the
PSCO is satisfied with the general impression, the inspection can be confined, and the
observed deficiency(ies) should be recorded.
Alternatively, if clear grounds regarding technical and managerial aspects of the
ship’s equipment and its crew or pollution prevention issues are found, the PSCO can
decide to conduct a more detailed inspection, which covers the ship’s construction,
equipment, manning, living and working conditions, and compliance with onboard
operational procedures. Similar to the initial inspection, any deficiency detected is
recorded. If a detainable deficiency (i.e., serious deficiency ground for detention)
is found, the ship can be detained. A ship can apply for a follow-­up inspection,
which is for verifying the rectification of deficiency(ies) found at the previous initial

* 
Here company refers to the ISM company
Table 2.1   Information sheet of NIR adopted by the Tokyo MoU

Parameters Values Weighting points Criteria for LRS

Ship type Chemical tanker, gas carrier, oil tanker, bulk carrier, 2
passenger ship, and container ship
Ship age (calculated based on All types with age >12 years 1
the keel laid date)
Flag performance in Black-­ Black 1 White and should be
Gray-­White list of Tokyo IMO audit
MoU
RO performance evaluated by Low/very low 1 High and should be an
Tokyo MoU RO recognized by the
Tokyo MoU
Company performance Low/very low/no inspection within previous 36 months 2 High
evaluated by Tokyo MoU
Deficiencies within previous 36 Inspections which recorded over five deficiencies The number of All inspections have five
months inspections which or less deficiencies,
recorded over five and the ship has at
deficiencies least one inspection
within previous 36
months
Detentions within previous 36 Three or more detentions 1 No detention
months
Source: Tokyo MoU [11].
14  Machine learning and data analytics for maritime studies

Table 2.2   Inspection time window of each SRP

SRP Criteria Inspection time window

HRS When the sum of weighting points ≥4 2–4 months
SRS Neither HRS nor LRS 5–8 months
LRS All the criteria for LRS are met 9–18 months
Source: Tokyo MoU [11].

inspection(s). It can be carried out onsite a ship or by remote mode (which is also
called remote follow-­up inspection). A general PSC inspection is not charged, while
port state charges can be expected in follow-­up inspections.

2.2.4 Inspection results
After a PSC inspection, inspection results, i.e., deficiency(ies) identified and the
detention decision are recorded in the relevant forms and reported in the cen-
tral database of the corresponding MoU. To be more specific, deficiency codes are
derived from major maritime conventions and regulations and are specified and
constantly updated by individual MoUs and might be slightly different among the
MoUs. Deficiency codes adopted by the Paris MoU (effective from 1 July 2021)
are shown in Table  2.3, and those adopted by the Tokyo MoU (effective from 2
December 2019) are shown in Table 2.4.
A ship is detained if the PSCO decides it is unseaworthy due to the identification
of serious deficiency(ies), while the captain has the right to appeal against a deten-
tion order. Although the general purpose of PSC is to prevent a substandard ship
from proceeding to sea, the detention of a ship is a serious matter involving many
issues, and thus PSCOs are cautious of the detention decision to avoid undue delays.

Table 2.3   List of deficiency codes in Paris MoU

Code Content Code Content

01 Certificates and documentation 02 Structural condition
03 Water/Weathertight condition 04 Emergency systems
05 Radio communication 06 Cargo operations including equipment
07 Fire safety 08 Alarms
09 Working and living conditions 10 Safety of navigation
11 Life-­saving appliances 12 Dangerous goods
13 Propulsion and auxiliary machinery 14 Pollution prevention
15 ISM 16 ISPS (the International Ship and Port
Facility Security Code)
18 MLC2006 99 Other
Source: Paris MoU [12].
 Ship inspection by port state control  15

Table 2.4   List of deficiency codes in Tokyo MoU

Code Item Code Item

01 Certificates and documentation 02 Structural conditions
03 Water/Weathertight condition 04 Emergency systems
05 Radio communications 06 Cargo operations including equipment
07 Fire safety 08 Alarms
09 Working and living conditions 10 Safety of navigation
11 Life-­saving appliances 12 Dangerous goods
13 Propulsion and auxiliary machinery 14 Pollution prevention
15 ISM 18 Labor conditions
99 Other
Source: Paris MoU [12].

Regarding the inspection information reported in central databases, some MoUs offer
rough inspection results, such as Caribbean MoU† and Riyadh MoU‡. More detailed ship
information together with the inspection results are provided by some other MoUs, such
as Paris MoU§, Tokyo MoU¶, Abuja MoU**, Black Sea MoU††, Indian Ocean MoU‡‡, and
Mediterranean MoU§§. Especially, statistics on inspections and their results are offered
by Paris MoU in graph user interface, which is more convenient and intuitive for the
users¶¶.

2.3 Data set used in this book

Data of the examples of algorithm illustration and code implementation are mainly
from PSC inspection records at the Hong Kong port. Unless otherwise stated, the
case data set used in this book is 1 991 initial inspection records from 2015 to 2017
at the Hong Kong port. The following fields are collected from the website operated
by the Tokyo MoU: inspection date, IMO number (IMO no.), ship type, keel laid
date, deadweight tonnage (DWT), gross tonnage (GT), SRP, classification society,
flag state, flag performance/RO/company performance given by the Tokyo MoU
(flag performance, RO performance, and company performance), last date of initial
inspection (last inspection date), number of deficiencies in last initial inspection (last
deficiency number), detention in last initial inspection (last detention), deficiency

† 
http://www.caribbeanmou.org/content/inspection-detention-data
‡ 
https://www.riyadhmou.org/basicsearch.html?lang=en
§ 
https://www.parismou.org/inspection-search/inspection-search
¶ 
https://www.tokyo-mou.org/inspections_detentions/psc_database.php
** 
http://www.abujamou.org/index.php?pid=125disclaimer
†† 
http://www.bsmou.org/database/inspections/
‡‡ 
https://www.iomou.org/HOMEPAGE/search_insp.php?l1=4&l2=26
§§ 
http://www.medmouic.org/Advanced
¶¶ 
https://www.parismou.org/inspection-search/inspections-results-kpis
Table 2.5   Five samples in the PSC case data set (part 1)

No. Inspection date IMO no. Ship type Keel laid date DWT GT SRP Classification society Flag state
1 December 29, 2017 9332626 Oil tanker February 2, 2008 74997 42893 SRP DNV GL AS Norway
2 December 27, 2017 9254484 Bulk carrier May 24, 2002 52364 30011 HRS Indian Register of Shipping India
3 December 27, 2017 9760500 Oil tanker October 2, 2015 106359 57164 SRS Lloyd’s Register Liberia
Table 2.6   Five samples in the PSC case data set (part 2)

Last Last
Flag RO Company inspection deficiency Last Deficiency Detainable
No. performance performance performance date number detention number Detention Deficiency codes deficiency codes
1 White High High September 0 No 4 No 01314/ 05118/ None
30, 2013 10106/ 07110*
2 White High Low November 9 No 14 Yes 03102/ 10118/ 11101/ 11101
19, 2015 03109/ 03107/
11101/ 11101/
11101/ 04102/
04116/ 14402/
03105/ 03108/
03105/ 07113 †
3 White High Medium None None None 0 No None None
*The nature of these codes is certificate and documentation - documents (SOPEP (The Shipboard Oil Pollution Emergency Plan), radio communications (operation of
GMDSS (Global Maritime Distress and Safety System) equipment), safety of navigation (compass correction log), and fire safety (fire fighting equipment and appli-
ances), respectively.
†
The nature of these codes is water/weathertight conditions (freeboard marks), safety of navigation (speed and distance indicator), water/weathertight conditions
(machinery space openings), water/weathertight conditions (doors), life-­saving appliances (lifeboats), emergency systems (emergency fire pump and its pipes), emer-
gency systems (means of communication between safety center and other control stations), pollution prevention - MARPOL annex IV (sewage treatment plant), water/
weathertight conditions (Covers [hatchway-, portable-, tarpaulins, etc.]), water/weathertight conditions (ventilators, air pipes, and casings), and fire safety (fire pumps
and its pipes), respectively.
18  Machine learning and data analytics for maritime studies

number of the current inspection (deficiency number), and detention of the current
inspection (detention). In addition, the specific deficiency codes and detainable defi-
ciency codes are also collected. Among them, “deficiency number” and “detention”
are potential target variables, and the others are potential input features. Three typi-
cal samples in the initial data set directly retrieved from the database are given in
Table 2.5 and Table 2.6 Table 2.6.
The three records above have their own characteristics. Record no. 1 is a com-
mon PSC inspection in the data set, which is an un-­detained inspection with last
inspection available. Record no. 2 is a detained inspection record with a large num-
ber of deficiencies detected (at 14) and detainable deficiencies detected. Record no.
3 is an un-­inspected ship by any port in the Tokyo MoU, and thus it has no historical
inspection information in the database.

References

[1] Brief history of IMO [online]. London: International organization (IMO).

Available from https://www.imo.org/en/About/HistoryOfIMO/Pages/Default.​
aspx [Accessed 9 Jan 2022].
[2] International convention for the safety of life at sea (SOLAS), 1974 [online].
London: International maritime organization (IMO). 2019. Available from htt-
ps://www.imo.org/en/About/Conventions/Pages/International-Convention-​
for-the-Safety-of-Life-at-Sea-(SOLAS)-1974.aspx [Accessed 12 Dec 2021].
[3] International convention for the prevention of pollution from ships (MARPOL)
[online]. London: International maritime organization. 2019. Available from
https://www.imo.org/en/About/Conventions/Pages/International-Convention-​
for-the-Prevention-of-Pollution-from-Ships-(MARPOL).aspx [Accessed 12 Dec
2021].
[4] International convention on standards of training, certification and watchkeep-
ing for seafarers (STCW) [online]. London: International Maritime Organization
(IMO). 2019. Available from https://www.imo.org/en/About/Conventions/​
Pages/International-Convention-on-Standards-of-Training-Certification-and-​
Watchkeeping-for-Seafarers-(STCW).aspx [Accessed 19 Jan 2022].
[5] International labour organization [online]. Geneva: Maritime labour conven-
tion. 2006. Available from https://www.ilo.org/global/standards/maritime-
labour-convention/lang--en/index.htm [Accessed 12 Feb 2022].
[6] Resolution A. 1155 (32) adopted on 15 december 2021 (agenda items 12 and
14) procedures for port state control, 2021 [online]. Geneva: International
Maritime Organization. 2022. Available from https://www.register-iri.com/​
wp-content/uploads/A.115532.pdf [Accessed 2 Feb 2022].
[7] Port state control [online]. London: International Maritime Organization
(IMO). Available from https://www.imo.org/en/OurWork/MSAS/Pages/​
PortStateControl.aspx [Accessed 25 Jul 2021].
[8] Knapp S. ‘The econometrics of maritime safety’. Recommendations to
Enhance Safety at Sea. 2007;96.
 Ship inspection by port state control  19

[9] Annual report [online]. Tokyo: Tokyo MOU. Available from http://www.
tokyo-mou.org/doc/ANN19-f.pdf [Accessed 25 Aug 2022].
[10] Selection of ships for inpsection [online]. London: Mediterranean MoU. 2014.
Available from http://www.medmou.org/Basic_Principlse.aspx#3 [Accessed
10 Dec 2020].
[11] Information sheet on the new inspection refime (NIR) [online]. Tokyo:
Tokyo MoU. 2014. Available from http://www.tokyo-mou.org/doc/NIR-
information%20sheet-r.pdf [Accessed 3 Mar 2022].
[12] List of paris mou deficiency codes [online]. Paris: Paris MOU. 2021. Available
from https://www.parismou.org/list-paris-mou-deficiency-codes [Accessed
15 Jan 2022].
[13] Tokyo MOU deficiency codes [online]. Tokyo: Tokyo MoU. 2019. Available
from http://www.tokyo-mou.org/publications/tokyo_mou_deficiency_codes.​
php [Accessed 15 Jan 2022].
This page intentionally left blank
 Chapter 3
Introduction to data-­driven models

This chapter aims to clarify the basic issues of data-­driven modeling and its applica-
tion in maritime transport. A predictive problem is first introduced, and the predic-
tive analysis in the field of maritime transport is then discussed with typical exam-
ples given. A comparison between the classic method and data-­driven modeling to
deal with prediction tasks is also provided. Finally, typical data-­driven modeling
approaches are briefly discussed.

3.1 Predictive problem and its application in maritime

transport

3.1.1 Introduction of predictive problem

Predictive problem is common in engineering and management. It aims to get the
correct output (also called target, label, and dependent variable) to an example
(also called instance and sample) based on the input (also called feature and inde-
pendent variable). Prediction can be realized in both qualitative and quantitative
manners. Qualitative prediction is mainly achieved by the judgment of practition-
ers based on their experience and expertise, while quantitative prediction is largely
dependent on historical data and mathematical tools. In the era of big data, practi-
cal prediction problems are more often addressed by quantitative methods, with
two classes of approaches widely used: theory-­based approach and data-­driven
approach.
Generally speaking, theory-­based modeling, as a classic approach, is derived
by certain system processes following physical laws and rules and is represented
by mathematical equations describing the whole process so as to estimate the target
based on system parameters. Alternatively, data-­driven modeling, as an emerging
approach which has received increasing attention, is based on data analytics while
emphasizing less on the underlying system processes. The target is estimated using
historical data based on a fitted predictive model, which can be a statistical model
or a machine learning (ML) model. The key part of data-­driven modeling is to find

1
The Hong Kong Polytechnic University, Hung Hom, Hong Kong, China
22  Machine learning and data analytics for maritime studies

connections between the system state variables, i.e., the input, internal, and output
variables.

3.1.2 Examples of predictive problem in maritime transport

In maritime transportation, prediction analysis is involved in many practical sce-
narios in both shipping and port operation and management, such as ship trajectory
prediction, ship navigation risk prediction, and safety management risk prediction
and safety management as a whole, which means to predict and reduce risks, ship
inspection planning, ship energy efficiency prediction, ocean freight market con-
dition prediction, ship destination and arrival time prediction, and port condition
prediction [1], to name just a few. We use the following examples to illustrate the
main parts of using theory-­based modeling and data-­driven modeling approaches to
address predictive problems.
Example 3.1: Ship fuel consumption prediction problem:
Ship fuel consumption prediction is a hot research topic in green shipping
management. In theory-­based modeling, models that strictly follow hydrody-
namics laws and physical principles of the vessel power system are developed.
Generally, a basic white-­box model starts from vessel calm water resistance and
additional resistance calculation, which is the basis of total resistance calcula-
tion. Then, the engine power required to drive the vessel to a certain speed under
such resistance, as well as the fuel consumption rate, is estimated. Typical meth-
ods include Holtrop-­Mennen method and Kristensen-­Lützen method [2]. As the
model structure and parameters are known because they are calibrated based on
physical and hydrodynamics laws, they can be applied in the initial stage of ship
design and sea trials without using historical sailing data. Alternatively, data-­
driven models largely rely on historical vessel sailing records, which can be col-
lected from ship noon reports and onboard sensors. A suitable regression model
for fuel consumption rate prediction is constructed using proper data and features
extracted from available historical sailing records, and its performance is then
validated. The representative of data-­driven model for ship fuel consumption pre-
diction is the so-­called “cubic law,” where a ship’s fuel consumption rate at sea is
regarded to be proportional to its sailing speed to a power of ‍˛ = 3‍. Other popu-
lar data-­driven models in existing literature include statistical modeling such as
multiple linear regression analysis and polynomial regression, and ML models
such as artificial neural networks (ANNs), tree-­based models, and support vector
regressor (SVR).
Example 3.2: Ship trajectory prediction problem:
Accurate ship trajectory prediction is the prerequisite for collision preven-
tion and sailing risk analysis. In theory-­based modeling, ship trajectory predic-
tion is mainly achieved by simulation, considering the vessel’s real-­time sailing
behavior, resistances, and hydrodynamic equations. Ship trajectory prediction
using data-­driven modeling approach is mainly powered by historical temporal
 Introduction to data-­driven models  23

trajectory points from vessel automatic identification systems data. These time-­
series data are then mined by classic ML models such as extreme learning
machine, support vector machine (SVM), and k-­nearest neighbors (KNNs), as
well as more advanced deep-­learning models like recurrent neural networks, so
as to figure out the sailing patterns that can be further used for vessel trajectory
prediction.

3.1.3 Comparison of theory-based modeling and data-driven

modeling
We summarize the advantages and disadvantages of theory-­based modeling and
data-­driven modeling based on the above two examples in Table 3.1.

3.1.4 Popular data-driven models

Two common approaches for data-­driven modeling are statistical modeling and ML
modeling. We introduce the definition, modeling process, examples, and when to
use the two approaches in this subsection.

3.1.4.1 Statistical modeling
A statistical model is a mathematical model representing the data generation pro-
cess based on a set of statistical assumptions. In statistical modeling, regression
analysis is a set of statistical processes with the aim to estimate the relationship
between a dependent variable (i.e., output) and a set of independent variables
(i.e., input) using a training set that is constituted by historical data, so as to pre-
dict the output of an unseen this means an example that is not used for construct-
ing and tuning the model example. After data acquisition, the statistical mod-
eling process begins from data preprocessing and feature extraction. As statistical
modeling relies on statistical assumptions, reasonable assumptions are made on
the data generation process, and then suitable regression models are selected.
Next, the parameters of the regression models are estimated using the historical
data. Finally, the fit and generalization abilities of the models are tested. Popular
statistical models for prediction tasks are linear regression, logistic regression,
polynomial regression, stepwise regression, ridge regression, lasso regression,
and elastic net regression. Given a dataset with ‍n‍samples and ‍m‍features denoted
by ‍D = f(xi , yi ), i = 1, ..., ng,‍ ‍xi 2 Rm ,‍ ‍yi 2 R ,‍ the basic forms and common parameter esti-
mation methods of these popular statistical models are shown in Table 3.2.

3.1.4.2 ML modeling
ML modeling belongs to the area of artificial intelligence, which aims to enable
a computer to think and act like humans and to solve complex tasks without or
with only little human intervention. A commonly used definition is given by Tom
Mitchell:
Table 3.1   Comparison between theory-­based modeling and data-­driven modeling

Quantitative
prediction approach Advantages Disadvantages
Theory-­based • System behavior and the prediction results are fully • The prediction accuracy can be low, which highly depends on
modeling tractable and interpretable. the assumptions made.
• No or few historical data of the system are needed, • Much a priori knowledge is needed.
and thus it can be applied in initial application stage • As the model structure and parameters are fixed, its
with low-­data acquisition costs. performance cannot be improved as data accumulate.
• It strictly complies with domain knowledge and thus
can be more acceptable by practitioners.
Data-­driven modeling • With proper data, feature, and model, its prediction • It is usually in a black-­box nature, and thus the model-­working
accuracy can be very high. mechanism and the predictions generated are highly likely to
• Little a priori knowledge is needed. be unexplainable.
• It can be constantly improved with the accumulation • Much historical data are needed for model construction, and
of data and refinement of model. the data acquisition costs can be very high.
• It can be very sensitive to data quality and quantity.
• The training data can be over-­fitted, leading to poor
generalization ability.
• The predictions might be contradictory to domain knowledge.
 Introduction to data-­driven models  25

Table 3.2   Popular statistical models

Common
parameter
estimation
Statistical model Basic form method
Linear regression ‍y = ˇ1 x1 + ... + ˇm xm + .‍ If ‍m = 1‍, it is simple Least squares
linear regression. Otherwise, it is multiple linear
regression.
Logistic regression Given input ‍x‍and model parameters ‍ ,‍ Weighted least
(for binary the probability of ‍y = 1‍is calculated by squares and
classification maximum
problems) T 1 likelihood
P(y = 1|x; ) =g( x) = Tx
.
‍ 1+e  ‍ estimation
Polynomial ‍y = ˇ 0 + ˇ 1 X + ˇ 2 X 2
+ ˇ 3 X3
+ ... + ˇ h X h
+  ,
‍ Least squares
regression where ‍ X ‍is the feature matrix of the training
set.
Stepwise regression Stepwise regression is used to address the Least squares
multicollinearity of variables in linear regression combined
models by automatically choosing predictive with F-­test
variables. There are three main approaches for and t-­test
stepwise regression.
• Forward selection: Starting from no variable
in the model and adding the variables from the
largest to the least contribution evaluated by F-­
test until no more variable(s) can be added.
• Backward elimination: Starting from all
candidate variables in the model and deleting
each variable whose loss gives the least
statistically significant deterioration of the model
fit evaluated by F-­test until no more variable(s)
can be deleted.
• Bidirectional elimination: A combination of
forward selection and backward elimination
to ensure adding the variable with the largest
contribution in each step, while all the added
variables in all runs are statistically significant
(otherwise, statistically insignificant variable(s)
will be deleted).
Ridge regression The regression function format is similar to linear Least squares
(regularized by an regression models. An L2 regularization term
L2 term) is included in the cost function to address the
over-­fitting problem as
Pn Pm 2
Pm 2
‍l = i=1 (yi  j=1 ˇj xij ) + j=1 ||ˇj ||2,‍
‍ > 0‍.
(Continues)
26  Machine learning and data analytics for maritime studies

Table 3.2  Continued

Common
parameter
estimation
Statistical model Basic form method
Lasso regression The regression function format is similar to linear Least squares
(regularized by an regression models. An L1 regularization term
L1 term) is included in the cost function to address the
over-­fitting problem as
Pn Pm 2
Pm
‍l = i=1 (yi  j=1 ˇj xij ) + j=1 ||ˇj ||1,‍
‍ > 0‍.

ML is a computer program that is said to learn from experience ‍ E ‍with respect to

some task ‍T ‍and some performance measure ‍ P ‍, if its performance on ‍T ‍, as measured
by ‍ P ‍, improves with experience ‍ E ‍.
ML models can be further divided into three categories, namely supervised
learning, unsupervised learning, and reinforcement learning. In supervised learn-
ing, a training set consisting of examples with features as well as targets is used
to train ML models, while in unsupervised learning, a training set containing
examples with features while without targets is used for ML model training.
Meanwhile, reinforcement learning aims to find suitable actions to be taken in a
given situation to maximize a reward. Among them, supervised learning is the
most widely used method to address predictive problems. Popular supervised
ML models include ANN, decision tree (DT), tree ensemble models, support
vector classifier (SVC)/SVR, and KNN. The basic idea and common forms of
these popular supervised ML models are shown in Table 3.3.
Although statistical modeling and ML modeling have different forms and
features, both are widely used to solve predictive models in multiple disciplines.
In summary, statistical modeling takes a form of a mathematical function, and
ML modeling uses computer algorithms to extract patterns from data without
relying on rule-­based programming not mentioned before; but this is how it
works. It is hard to compare their advantages and disadvantages. Nevertheless,
they have their own suitable application scenarios. Generally speaking, statisti-
cal models are more suitable to be applied when the input and output variables
take a linear or near-­linear form. In addition, as they are dependent on statistical
theory, they are more suitable to be applied when high-­model interpretability,
as well as the model and variable significant tests, are required. Meanwhile, ML
models should be adopted when the requirement of model accuracy is high or
when the data size is large and with many uncertainties. However, it should also
be noted that there is no rule of thumb in deciding which approach is more suit-
able to address a specific prediction problem. Instead, the one that can best meet
the actual needs should be found in a trial-­and-­error manner.
 Introduction to data-­driven models  27

Table 3.3   Popular supervised ML models

Model Basic idea Common forms

ANN An ANN model aims to extract linear Back-­propagation neural network,
combinations of the input features feed forward neural network,
so as to derive more effective radial basis network, and deep
features, and then calculate the neural networks.
target using a nonlinear function
of these derived features. An ANN
model usually consists of an input
layer, a hidden layer, and an output
layer.
DT The construction process of a DT Popular DT algorithms include
mainly involves consecutive node ID3 (short for iterative
splitting from the root node to the dichotomizer 3) and C4.5
leave nodes, aiming to minimize (an extension of ID3) for
the loss function. Each split classification tasks, and CART
requires selecting a feature together (standing for classification
with one of its values. and regression tree) for both
classification and regression
tasks.
Tree ensemble To reduce the tendency of over-­fit Random forest, extremely
models and local optimization of a single randomized trees, adaptive
DT and thus to improve model booster, gradient boost, and
generalization ability, tree ensemble extreme gradient boosting.
models are developed by applying
bagging (bootstrap aggregating) or
boosting on a set of DTs.
SVC/SVR As the names suggest, SVC is for Nonlinearity in SVM is achieved
classification task, and SVM is by kernel function. Popular
for regression task, while both kernel functions include
methods belong to the class of Gaussian kernel, Gaussian
SVM. An SVC aims to produce a kernel radial basis function,
nonlinear boundary (hyperplane) sigmoid kernel, and polynomial
that maximizes the distance margin kernel.
between two classes, and an SVR
aims to minimize the generalization
bound by developing a linear
regression function in a high-­
dimensional feature space.
KNN KNN is memory-­based with no model Distance metric is an important
to be fit. For a given point to be issue affecting the performance
predicted, the k training points of KNN algorithm. Popular
nearest to the point of concern distance metrics are Euclidean
are first identified. Then, for distance, Minkowski distance,
classification problem, the output is Manhattan distance, Cosine
predicted by majority vote among distance, Jaccard distance, and
the k neighbors; for regression Hamming distance.
problem, the predicted target of the
point is the average output of the k
neighbors.
28  Machine learning and data analytics for maritime studies

References

[1] Yan R., Wang S., Zhen L., Laporte G. ‘Emerging approaches applied to mari-
time transport research: past and future’. Communications in Transportation
Research. 2021;1:100011.
[2] Yan R., Wang S., Psaraftis H.N. ‘Data analytics for fuel consumption man-
agement in maritime transportation: status and perspectives’. Transportation
Research Part E. 2021;155:102489.
 Chapter 4
Key elements of data-­driven models

This chapter aims to thoroughly introduce and discuss the major issues of data-­driven
models, with a focus on machine learning (ML) models and their construction procedure.
We first compare three popular data-­driven models, namely, statistical model, ML model,
and deep learning (DL) model. Then, the whole procedure of developing a data-­driven
model will be provided. Key elements from various aspects of the whole process will be
covered in detail.

4.1 Comparison of three popular data-driven models

We first compare the differences among the three popular data-­driven models,
namely, statistical model, ML model, and DL model from various perspectives as
presented in Table 4.1. It is noted that although DL is a type of ML as DL models
are based on deep neural networks (DNNs) consisting of multiple hidden layers, we
treat them separately as they have several major differences from other ML models.
In the remainder of this book, our main focus will be on using ML models to address
practical problems in maritime transport.

4.2 Procedure of developing ML models to address maritime

transport problems
The whole procedure of developing an ML model to address a specific practical
problem in maritime transport is shown in Figure 4.1. It starts by specifying the
problem to be solved, which needs extensive communications and discussions with
the parties of interest. The main goal is to understand the business scenarios where
the ML model will be applied and the expected outcomes of the stakeholders. Then,
the feasibility of addressing the problem should be evaluated considering the avail-
able data, technology, and human resources. If needed, the requirements and out-
comes can be adjusted after reaching an agreement among the parties. Then, data
needed to fulfill the requirements and outcome are to be collected from available
(and also may be various) resources. After that, feature engineering is conducted to
enable the available data to better describe the essence of the problem studied, so
as to improve the prediction accuracy. Next, ML models are developed and are then
evaluated, assessed, and refined. It should be noted that although the basic sequence
Table 4.1   Comparison of statistical, ML, and DL models

Data and computing Prediction accuracy

Model name Basic idea power requirement level Explainability Typical applications
Statistical A statistical model is based Structured data, low Medium, highly Explainable like a Time-­series analysis,
models on statistical learning computing power depending on “white box” attribution analysis,
theory, with the goal required whether reasonable survival analysis
of finding a predictive assumptions are
function in an explicit made and proper
form by leveraging structures are
historical data adopted
ML models An ML model is constructed Mainly structured data, Medium to high, Most are un-­explainable Medical and healthcare
to make a computer to be medium computing depending on data like a “black box”, management, fraud
able to perform tasks with power required characteristics. but some are detection, product
the process of “learning” Generally, ensemble explainable, such as recommendation,
while without explicit models have higher linear regression and dynamic pricing, and
“programming” predictive accuracy decision tree social network mining
DL models A DL model aims to mimic Structured and Medium to high, Un-­explainable like a Automated driving,
the behavior of human unstructured data, depending on data “black box” sentimental analysis,
brain (regarding structures high computing quantity and the natural language
and functions) and is power required model structure processing, digital
mainly achieved by [where GPU image processing,
constructing multi-­layer (graphics processing video and audio
DNNs unit) is usually processing, and
needed] automatic game
playing
 Key elements of data-­driven models  31

Figure 4.1  Procedure of developing ML models for maritime transport

problems*

Note that the basic sequence of the four steps in the rounded rectangle is data collection → feature
* 

engineering → model construction → model evaluation, selection and refinement. In practice, some of
them may be repeatedly operated to satisfy the requirements.
32  Machine learning and data analytics for maritime studies

of these four steps (from data collection to model evaluation, assessment, and refine-
ment) is generally like this, they might also be alternating and repetitive considering
the gap between the requirements as well as the expected outcomes and the devel-
oped ML models on hand. Finally, prediction results and model explanations are
obtained, and managerial insights are derived for improving the operation efficiency.
In the next subsections, we will discuss the key elements of the procedure in detail.

4.2.1   Problem specification

The problem specification step aims to understand the business problem to be
solved and then turn the business problem into a data science problem, which is
mainly achieved by communicating with and interviewing the parties of interest
and understanding the problem backgrounds. In this process, the following three
aspects should be explored, namely, objectives, resources, and requirements. To be
more specific, objectives mainly refer to the prediction targets of the problem and
the problem’s specific form. For example, to predict ship risk in PSC inspection,
the prediction target can be ship deficiency number or detention rate, among other
prediction targets. Resources include the available data sources and computational
power, especially when big data, real-­time or near-­real-­time prediction, and complex
prediction models are needed. Requirements regarding prediction accuracy, presen-
tation form, and model interpretability and explainability can be varied considering
different users and application scenarios. For example, academic and technical staff
may expect more details of the fundamental theories and the detailed implementa-
tion procedure of the prediction model, while concise and user-­friendly graphical
user interface might be more likely to be expected by onboard crew. Based on the
information gathered, the problem is then broken down into several smaller prob-
lems, so as to shed more light on the data to be collected and the models/methods
to be adopted.

4.2.2   Feasibility assessment

Feasibility assessment is usually conducted after the requirements are specified,
and the data to be collected and methodology to be adopted are planned. It aims to
determine whether the requirements of the project can be met given the budget, the
schedule, the anticipated technology, and the human resources. For any infeasibil-
ity, requirements might be updated or the technologies adopted should be modified
based on mutual agreements between the stakeholders and the data scientists.

4.2.3   Data collection

Data can be regarded as the most important component in data-­driven modeling, as
the quality and quantity of data used could have a significant impact on model accu-
racy and its generalization ability. In this step, required data are collected and fused
from multiple data sources. The data collected can be structured and unstructured
data. Especially, structured data are usually in a quantitative and formatted form,
which are mainly represented by numbers and strings. In contrast, unstructured data
are usually in a qualitative form and can be text files, audio and video files, and
 Key elements of data-­driven models  33

figures. We briefly introduce common data sources that are popular in maritime
transport research as follows.

4.2.3.1   Automatic identification system

Automatic identification system (AIS) is required to be equipped on vessels involv-
ing international voyages with a gross tonnage above 300 and on all passenger ships.
It was first developed in the 1990s as a short-­range identification and tracking sys-
tem for the purpose of collision avoidance. It then has a wider application in, e.g.,
fishing fleet monitoring and control, vessel traffic management, maritime security,
aids to navigation, search and rescue, accident investigation, and fleet and cargo
tracking. Vessel dynamic information (geographical location, sailing speed, naviga-
tion status, course, heading degree, and rate of turn), static information [ship name,
type, size, draught, maritime mobile service identity (MMSI), IMO number, and
call sign], and voyage information (destination and the estimated time of arrival) are
given in AIS reports. The update interval of dynamic AIS reports, ranging from 2 to
10 seconds, highly depends on the vessel sailing speed, and the static and voyage
AIS reports are updated every 6 minutes or when any field contained by the static
and voyage AIS reports is changed. For a detailed introduction to AIS, readers are
referred to References 1 and 2.

4.2.3.2   Ship specifications

Ship specification information, including but not limited to basic ship characteris-
tics [current and former vessel names, vessel identification numbers (IMO number,
MMSI, and call sign), vessel type, year of built and builder, tonnages and dimen-
sions, and engine information], ship operation information (current and former
flag states, management companies, operators, and owners), ship survey and main-
tenance records, and PSC detention records, can be found in commercial vessel
databases such as Lloyd’s List Intelligence*, Shipping Intelligence Network†, and
MarineTraffic‡. Features of ships with certain characteristics, e.g., in certain clas-
sification societies (e.g., American Bureau of Shipping§) and the sunken ships (e.g.,
International Registry of Sunken Ships¶) can also be found in specific databases.

4.2.3.3   Ship sailing records

Ship sailing records can come from automated onboard sensors based on the Internet
of Things technology and the manually filled ship noon report. Common sensors
on ships to monitor their performance are GPS receiver, speed log, revolutions per
minute and torque meter, shaft power tester, fuel flow meter, wind anemometer, and

* 
https://ihsmarkit.com/products/maritime-ships-register.html
† 
https://www.clarksons.net/portal/
‡ 
https://www.marinetraffic.com/en/ais/home/centerx:-12.0/centery:25.0/zoom:4
§ 
https://ww2.eagle.org/en.html
¶ 
http://www.shipwreckregistry.com/
34  Machine learning and data analytics for maritime studies

water depth sensor. Data are automatically collected and then stored in the vessel’s
central database and transmitted to port authorities.

4.2.3.4   Ship accident data

Public ship accident and incident data around the world can be found in the Global
Integrated Shipping Information System (GISIS) provided by the IMO**. Factual
data collected from various sources and more elaborated information extracted from
the reports of investigation of casualties received by the IMO are provided by the
GISIS. Ship casualties are classified into “very serious casualties,” “serious casual-
ties,” “less serious casualties,” and “marine incidents.” Regional marine accident
reports and statistics are also provided by local governments and departments. For
example, the Marine Department of Hong Kong established the Marine Accident
Investigation Section, which is responsible to investigate all marine accidents occur-
ring in Hong Kong and onboard Hong Kong registered ships and to conduct official
inquiries under certain ordinances. Reports and summaries of accident investigation
of individual ships carried out by the Marine Accident Investigation Section are
published on the Marine Department’s website††.

4.2.3.5   Port statistics

Key performance indicators are published by some ports that we refer to as port sta-
tistics. Common fields are visiting vessel statistics, port container throughput, port
cargo throughput, marine accidents, and bunker sales. Fields published by different
ports might differ from each other. For example, the Hong Kong Port‡‡ publishes
monthly, quarterly, and yearly vessel statistics, port container and cargo throughput,
and cross-­boundary ferry terminal statistics.
In addition to the common databases discussed so far, structured and unstruc-
tured data can be collected via other means, such as surveys, interviews, question-
naires, and observations.

4.2.4   Feature engineering

Feature engineering is the most time-­consuming and important step in the pro-
cedure of ML model development. Feature engineering aims to prepare proper
input data that are compatible with an ML model, so as to improve its perfor-
mance. Feature engineering is a huge project, and we only cover its main parts
in this subsection concisely from the following aspects: data cleaning, feature
extraction, feature pre-­processing, feature encoding and scaling, and feature
selection.

** 
https://gisis.imo.org/Public/MCI/Default.aspx
†† 
https://www.mardep.gov.hk/en/publication/publications/reports/ereport.html
‡‡ 
https://www.mardep.gov.hk/en/fact/portstat.html
 Key elements of data-­driven models  35

4.2.4.1   Data cleaning

After the raw data are collected, data quality should be (initially) checked (and will
then be carefully checked and improved after feature extraction) considering their
integrity, uniqueness, authoritativeness, legitimacy, and consistency. The explana-
tion of the above factors using PSC data as an example and the corresponding solu-
tions are given in Table 4.2. In addition to the above factors, some details also need
attention. For example, numbers stored as strings should be converted to numeric
types; extra spaces in data should be deleted; and the format of timestamp should be
converted from string format to time format.

4.2.4.2   Feature extraction

Feature extraction aims to create new features from the existing ones and delete the
original ones to make the features more informative and concise. Feature extraction
is one approach for dimension reduction (and the other is feature selection which
will be covered later in subsection 4.2.4.5), which refers to a process of reducing
the number of features in the data set while keeping as much variation in the origi-
nal data set as possible. Feature extraction methods can be linear and nonlinear.
We introduce principal components analysis (PCA) and linear discriminant analysis
(LDA) as the representatives of linear methods, and kernel PCA and multidimen-
sional scaling (MDS) as the representatives of nonlinear methods.
PCA: PCA is an unsupervised linear transformation to project the original data
to lower-­dimension orthogonal directions of high variance, aiming at using less fea-
tures in lower dimensions (i.e., principal components) to represent the original fea-
tures in higher dimensions. The orthogonality property results in very low or almost
close to zero correlation in the projected data.
LDA: LDA is a supervised learning-­based feature extraction method and also
an ML classifier. It aims to maximize the distance between the mean of each class
and minimize the spreading within each class. The idea is in line with that of clas-
sification, and thus it can be expected that LDA can lead to better performance in
classification ML models.
Kernel PCA: The assumption of PCA is that there is a linear hyperplane for data
projection. However, when a data set is not linearly separable, linear PCA might not
be a good choice. To address this issue, kernel PCA using kernel function to project
data into a higher dimensional feature space is proposed, where the nonlinear data
are linearly separable in the kernel space.
MDS: MDS has two major types: metric MDS and non-­metric MDS, where
the former attempts to calculate the similarity/dissimilarity of data measured by
distances based on geometric coordinates, and the latter is designed to deal with
ordinal data. In general, dissimilarities of each pair of points in the original high-­
dimensional space are first calculated, and then the points are mapped to lower-­
dimensional space while the dissimilarities of the point pairs are preserved as much
as possible.
Table 4.2   Explanation and solution of data cleaning

Factor Explanation Example in PSC data Solution

Integrity Data integrity is the overall Ship age is missing in the data The missing features can be added by
accuracy, completeness, and set. searching for more databases or
consistency of data. calculating from other features.
Uniqueness Data uniqueness means that a One ship inspection record is Only one of the duplicate records is kept.
single record only appears once recorded more than once in
in the data set. the data set.
Authoritativeness Data authoritativeness requires that The number of deficiencies Data from authoritative sources should be
the data should be legislated and of a ship in one inspection used.
regulated, and should come from recalled by the captain is
authoritative source or reference different from that searched
that is widely recognized by on the MoU’s official
experts in the field or industry. website.
Legitimacy Data legitimacy means that The age of an inspected ship is Rules considering ground truth or based on
data values should comply 1,000 years. domain knowledge are set to specify the
with ground truth and expert data range (ways to process outliers are
knowledge. covered in the following steps).
Consistency Data consistency requires that the The format of inspection date Data format should be unified.
same type of data appears in is varied, e.g., yyyy-­mm-­dd,
different places throughout the dd-­mm-­yyyy, and mm-­dd-­
data set to be matched. yyyy.
 Key elements of data-­driven models  37

Table 4.3   Methods to handle missing values in data sets

Method Description Note

Search the missing Query different sources and Data format of external sources
values from databases to search for specific should be in compliance with
various sources missing values. the original data format.
and databases
Fill with a global Fill missing values of all features The assumption of this method is
constant using a global constant, such that null value is informative
as “−1,” “0,” and “N.A.” in itself, that is, it is indeed
different from valid values.
Although this method is
intuitive and easy to operate,
it cannot be used or will cause
confusion sometimes.
Fill with statistics of Fill missing values of a certain Record similarity can be
the feature feature using its mean, median, evaluated by k-­means
or mode over all or similar clustering and distance
record(s) in the data set. measures. An extra binary
feature indicating whether each
record has a missing value of a
certain feature can be added.
Fill using the Use the feature value of the next N.A.
forward and or previous record (or the
backward method mean feature value of the next
or previous few records) in
time-­series data or a quantity
that is stable over a period
of time (e.g., ship draft value
over a voyage).
Fill with regression-­ Predict the missing value of a One drawback is that the
based approach certain feature using other correlation between features
features based on regression can increase.
models.

4.2.4.3   Feature pre-processing

Feature pre-­processing aims to deal with missing values and outliers in features.
Missing values cannot be simply ignored or deleted in the original data set; instead,
they must be carefully handled. One exception is dealing with a record (i.e., row data)
or a feature (i.e., column data) where most data fields are missing. Consequently, the
record or feature is meaningless. In other cases, there are several ways to handle
missing data in features, which we list in Table 4.3.
An outlier is an observation that is far away in a random sample from the data
set. Outliers can have a significant impact on the assumptions of data distribution
and the prediction models constructed, and thus they need to be detected and prop-
erly treated. Outliers can be univariate and multivariate, where the former is an
extreme value of a specific feature, whereas the latter is a combination of values that
are not likely to occur together with an observation. For example, if there is a PSC
38  Machine learning and data analytics for maritime studies

inspection record saying that ship type is “handymax carriers” and the dead weight
tonnage is 80,000, it is an outlier, as the dead weight tonnage of ship type “handy-
max carriers” is between 40,000 and 60,000. We only discuss univariate outliers
in this subsection. Common approaches of univariate outlier detection are listed in
Table 4.4.
After an outlier is detected, whether to drop the outlier needs careful consider-
ation. Some outliers are detected due to inherent variations in feature values, and
they are normal values instead of noise. For example, ship age 42 is detected as an
outlier in the PSC data set used in Reference 3. Nevertheless, the oldest ship still
afloat in the world is more than 200 years old, and thus the 42-­year-­old ship should
not be regarded as an outlier. Actually, given all other conditions equal, older ships
are usually associated with a larger number of deficiencies and a higher detention
probability and is thus the focus on PSC. Therefore, this outlier detected should
not be deleted or modified. Actually, the above example is just one case where the
detected outlier should be kept. When there are lots of outliers (e.g., more than 30%)
or the outliers can be rectified by capping, assigning new values, and data transform-
ing, they should not be dropped. In some other cases, the outliers should be dropped,
e.g., when it is sure that they are wrong, there are many data while outliers are few,
and they come from exceptional samples that should not be included.

4.2.4.4   Feature encoding and scaling

Feature encoding is needed for categorical features, which are non-­numeric and
are divided into groups and categories, with the aim to turn categorical data into
numerical data that are compatible with most ML models. Meanwhile, feature scal-
ing is applied to numerical features with varying scales when certain ML models are
adopted. Their details will be introduced in the following paragraphs.
There are basically three types of categorical features: binary feature (e.g., yes
and no; true and false), ordinal feature (e.g., low, medium, and high; bad, average,
and good), and nominal feature (e.g., container ship, bulk carrier, and passenger
ship). Encoding a binary feature is the easiest: just encode positive samples to 1
and negative samples to 0. For ordinal feature and nominal feature, we summa-
rize the ideas, examples, and application scenarios of popular encoding methods in
Table 4.5 that are widely applicable to PSC-­related research.
In addition to the common feature encoding methods shown in Table 4.5, there
are some other popular feature encoding methods that can be adopted to address
other practical problems in maritime transportation. One is hash encoding, which
aims to convert categorical values into numerical values using hashing functions
and is applicable to nominal feature. Another is bin counting, which is based on sta-
tistics that represent the relationship between the categorical feature and the target,
where the relationship is measured by a conditional probability of the target under a
given value of the categorical variable and is applicable to both nominal and ordinal
features.
Feature scaling should be applied in developing ML models based on geometric
distance or with weights assigned to features, such as linear regression (when it is
Table 4.4   Methods to handle outliers in data sets

No. Method Description

1 Method based on standard deviation If a feature value has a distance to the average higher than a given threshold (usually between 2 and
4) of the standard deviation, it is regarded as an outlier.
2 Method based on percentiles Assume a certain percentage of feature values from the top or from the bottom as outliers
considering data distribution and domain knowledge.
3 Interquartile range (IQR) Sort the values of a feature in ascending order and split them into four equal parts by Q1
‍ ‍, Q2
‍ ‍, and Q3
‍ ‍
(i.e., the 25th, 50th, and 75th percentile of data). ‍IQR = Q3  Q1.‍ Values below Q1‍  1.5  IQR‍
and above Q3 ‍ + 1.5  IQR‍are regarded as outliers.
4 Box plot Box plot can demonstrate the data visually, which is a data distribution figure with median, quartiles
of Q3
‍ ‍and Q1 ‍ ‍, lower bounds calculated by Q1‍  1.5  IQR‍and Q1 ‍  3  IQR‍and upper
bounds calculated by Q3 ‍ + 1.5  IQR‍and Q3 ‍ + 3  IQR‍shown. It should be noted that which
records are regarded to be outliers is highly dependent on subjective judgment.
 40  Machine learning and data analytics for maritime studies

Table 4.5   Methods of feature encoding in data sets

Encoding Application
method Idea Example scenarios
One-­hot Convert each Feature “ship type” has three Nominal feature.
encoding category value values: “container,” “bulk Note that although
into a new carrier,” and “passenger ship.” it can also be
column and We first convert this feature applied to ordinal
assign “1” or into three features: “is_ features, the order
“0” to the new container,” “is_bulk_carrier,” cannot be retained.
columns, where and “is_passenger_ship” and
“1” is assigned encode “container” as “(1, 0,
to the column of 0),” “bulk carrier” as “(0, 1,
the category the 0),” and “passenger ship” as
sample belongs “(0, 0, 1)” where each digit is
to, and “0,” for one new feature.
otherwise.
Target Convert each value Feature “ship age” has three Ordinal features
encoding in a column values: “young,” “middle-­
to a number aged,” and “old.” We encode
considering their “young,” “middle-­aged,” and
inherent orders. “old” to “1,” “2,” and “3.”

solved by gradient descent algorithm), ‍k ‍-­means, ‍k ‍ nearest neighbors, and PCA. In

contrast, feature scaling is not needed in ML models that are not distance-­based
and no weights are assigned, such as Naive Bayes and tree-­based models. There
are mainly three types of feature scaling methods, which are presented as follows.
Suppose that we have a set of values of a feature denoted by fx ‍ 1 , ..., xn g‍:
xi  min(x1 , ..., xn )
Min–max scaling: xi = 0 , where ‍x0i ‍is the feature
‍ max(x1 , ..., xn )  min(x1 , ..., xn ) ‍
value after min–max scaling, ‍min(x1 , x2 , ..., xn )‍ and ‍max(x1 , x2 , ..., xn )‍ are the mini-
mum and the maximum values of all the feature’s values, respectively. The range
after min–max scaling is [‍ 0, 1].‍
x  mean(x1 , ..., xn )
Mean normalization: x0i = , where ‍mean(x1 , x2 , ..., xn )
‍ max(x 1 , ..., xn )  min(x1 , ..., xn ) ‍
mean(x1 , x2 , ..., xn )‍is the mean of the feature’s values.
x  mean(x1 , ..., xn )
Standardization: x0i = , where std(x
‍ 1 , x2 , ..., xn )‍ is the stan-
‍ std(x1 , ..., xn ) ‍
dard deviation of the feature. It makes the feature values have zero-­mean and unit
variance.

4.2.4.5   Feature selection

Feature selection is a critical problem in feature engineering. It aims to delete irrel-
evant and redundant features and find the optimal feature subset (combination of
features), so as to improve model performance and reduce computation time. Both
 Key elements of data-­driven models  41

supervised and unsupervised methods can be used for feature selection. Supervised
feature selection techniques use the prediction target, and it can be further divided
into three categories: filter, wrapper, and embedded, while unsupervised feature
selection techniques ignore the target. We present common approaches for super-
vised feature selection as follows. Readers are referred to [4] for more information
on unsupervised feature selection techniques.

4.2.4.6   Pearson’s correlation coefficient (Pearson’s ‍r ‍)

Pearson’s ‍ r ‍ is a supervised method based on the filter that measures the linear cor-
relation between two numerical variables. Pearson’s ‍r ‍between feature ‍x ‍and target
cov(x, y)
‍y ‍ is calculated by x,y = , where ‍cov(x, y)‍ is the covariance of feature ‍x ‍
and target ‍y ,‍ and ‍‍x ‍ and ‍y ‍ are
x y ‍
the standard deviations of feature ‍x ‍ and target ‍y ,‍
respectively. ‍x,y 2 [1, 1],‍ with ‍1‍(‍ +1‍) indicates complete negative(positive) cor-
relation and ‍0‍ indicates no linear correlation. The basic idea of using Pearson’s ‍r ‍
for feature selection is first to calculate Pearson’s ‍ r ‍between each of the numerical
features and the numerical target. Then, a certain number of features with the highest
absolute Pearson’s ‍r ‍s or features with absolute Pearson’s ‍r ‍above a given threshold
are selected as the optimal feature subset.

4.2.4.7   Spearman’s rank coefficient (Spearman’s ρ)

Spearman’s ρ is a supervised method based on the filter that measures the nonlinear
correlation between two numerical variables. Spearman’s ρ between feature ‍x ‍ and
cov(R(x), R(y))
target ‍y ‍ is calculated by R(x),R(y) = , where ‍cov(R(x), R(y))‍ is the
‍ R(x) R(y) ‍
covariance of the rank variables, and ‍R(x)‍ and ‍R(y)‍ are the standard deviations of
the rank variable of feature ‍x ‍and target ‍y ‍. The basic idea of using Spearman’s ρ for
feature selection is first to calculate Spearman’s ρ between each of the numerical
features and the numerical target. Then, certain number of features with the highest
absolute Spearman’s ρ or features with absolute Spearman’s ρ above a given thresh-
old are selected as the optimal feature subset.

4.2.4.8   Chi-squared test (

‍ ‍test)
2
2
‍ ‍ test is a supervised method based on the filter that measures the correlation

between two categorical variables (one is a categorical feature and the other is the
categorical target). Suppose that there are ‍n‍ observations belonging to ‍ K ‍ catego-
ries and category ‍k ‍has mk samples. A null hypothesis gives the probability ‍pk = K1 ,‍
meaning that a sample falls into category ‍k ‍ and the ‍k ‍ categories are independent.
P (mk  npk )2
Then, 2 = Kk=1 . The basic idea of using  2
‍ ‍ test for feature selection
‍ npk ‍
42  Machine learning and data analytics for maritime studies

is first to calculate  2
‍ ‍test between each of the categorical features and the categori-
cal target. Then, a certain number of features with the highest  2
‍ ‍statistic or features
with  2 statistic above a given threshold are selected as the final feature subset.
‍ ‍

4.2.4.9   Mutual information

Mutual information (MI) is a supervised method based on filter that measures the corre-
P P p(x, y)
lation between two categorical variables. MI(x, y) = x2X y2Y p(x, y) log ,
‍ p(x)p(y) ‍
where ‍ X ‍ and ‍ Y ‍are the domains of ‍x ‍and ‍y ,‍ ‍p(x, y)‍is the joint probability of ‍X = x ‍
and Y‍ = y ,‍ and ‍p(x)(‍ ‍p(y))‍ is the probability of ‍X = x ‍(Y‍ = y ).
‍ The basic idea of using
MI for feature selection is first to calculate MI between each of the categorical fea-
tures and the categorical target. Then, a certain number of features with the highest
MI or features with MI above a given threshold are selected as the optimal feature
subset.

4.2.4.10   Recursive feature elimination

Recursive feature elimination (RFE) is a supervised method based on the wrapper
that removes unimportant features iteratively based on the evaluation metric. An ML
model that assigns weights to features is trained on the entire initial set with all fea-
tures. Then, the least important feature is removed from the current feature set, and
a new ML model is trained using the remaining features. The above step is repeated
until the desired number of features is selected as the optimal feature subset.

4.2.4.11   Feature selection based on regularization

This feature selection method belongs to the embedded method, which is usually
integrated as a part of an ML algorithm. Regularization can shrink the coefficients
(i.e., weights) of the unimportant/low impact features toward zero, and thus achieve
the goal of feature selection. Common approaches include least absolute shrinkage
and selection operator (LASSO), ridge, and ElasticNet where L1 regularization, L2
regularization, and the combination of L1 and L2 regularization are used, respec-
tively. These three approaches are introduced in Table 4.2 of section 4.2.4.1.

4.2.4.12   Feature selection based on feature importance

This feature selection method belongs to the embedded method and is usually
achieved by tree-­based models and permutation. In tree-­based models, such as deci-
sion tree, random forest, gradient boosting decision tree, XGBoost, and LightGBM,
feature importance is usually estimated by the impurity reduction in node split pro-
cess using the selected features, and features leading to higher impurity reduction
should be more important. In permutation-­based feature selection, an initial ML
model with original features is first trained. Then, the features are permuted one by
one to form new data set for model training, and the increase in the prediction error
is observed. Feature leading to a higher increase in error should be more important.
 Key elements of data-­driven models  43

4.2.5   Model construction

The development of ML models is one of the most crucial modeling steps. One
critical issue is that as there are many algorithms to develop different ML models to
choose from, how to choose a proper one? Another critical issue is how to tune the
hyperparameter values so as to improve the ML model’s performance? These two
issues are discussed in this subsection.

4.2.5.1 Model evaluation metrics

Different metrics should be adopted to evaluate regression and classification ML
models. For regression models where the output is a numeric value, common met-
rics are mean-­squared error (MSE), root-­mean-­squared error (RMSE), mean abso-
lute error (MAE), mean absolute percentage error (MAPE, where none of the output
values can be zero), and r-­squared (‍ R2‍). Suppose that there is a test set with ‍n‍sam-
ples where the real output and predicted output of each sample i‍ ‍ are denoted by yi
and ‍yOi ‍, i‍ = 1, ..., n‍. The metrics are calculated as follows:
1X
n

MSE = (yi  yOi )2 ,

n i=1
v
u n
p u1 X
RMSE = MSE = t (yi  yOi )2 ,
n i=1
1X
n

MAE = jyi  yOi j,

n i=1
100% X yi  yOi
n

MAPE = j j,
n yi
Pn i=1
1X
n
(yi  yOi )2
R2 = 1  Pi=1
n , where y
N = yi .
‍ i=1 (yi  y
N )2 n i=1 ‍

For the metrics mentioned above, MSE is the most widely used metric for regression
problems as it is older and well-­established. It is also differentiable, and thus there
are many good algorithms that can be applied to optimize loss functions based on
MSE. However, the unit of MSE is the square of the original data unit, and hence
this makes it hard to be interpreted. To remedy this issue, RMSE is proposed where
the unit is in compliance with the original data unit. Both metrics can be very sensi-
tive to the outliers in the test set, especially for the MSE where the square operation
is applied. This problem can be largely addressed by the MAE, whose unit also com-
plies with the original data and the metric is not that sensitive to outliers compared
to MSE. However, as it contains an absolute function that is not differentiable, the
optimization process is difficult. MAPE is a relative error rate that also considers the
original output, but the original output values cannot be zero. ‍ R2‍is also widely used
in statistics analytics, which represents the proportion of the variance in the output
that can be explained by the features. It can be seen that the meanings as well as the
pros and cons of the metrics for regression problems are different from each other.
44  Machine learning and data analytics for maritime studies

Table 4.6   Confusion matrix for a binary classification problem

Predicted value/Actual value Positive Negative

Positive True positive (TP) False positive (FP)
Negative False negative (FN) True negative (TN)

To evaluate ML models developed to address practical problems, it is recommended

that more than one metric should be used, and the metrics should comply with the
requirements.
For classification models where the output is a categorical variable, the situation
can be much more complex, which is mainly because the problem can be binary, i.e.,
the target can take two values (which are also called positive and negative classes;
e.g., detained or not detained), or multiclass, i.e., the target can take more than two
values (e.g., ship type of container, bulk carrier, and passenger ship). As multiclass
prediction problems can be turned into binary class prediction problems, we only
introduce the metrics for binary classification problem evaluation.
The metrics used for binary classification problems are based on the confusion
matrix as shown in Table 4.6.

4.2.5.1.1  Accuracy
Accuracy is a classic and intuitive classification metric, which measures the ratio of
the samples that are correctly classified to all the samples in the test set calculated
as follows:
TP + TN
Accuracy = .
‍ TP + FP + FN + TN ‍
Accuracy is suitable to be applied in balanced classification problems where the
numbers of samples in the positive class and the negative class are similar. In case
that the problem is imbalanced, where the positive class is very sparse, always pre-
dicting the target to be negative will get a very high accuracy.
Example 1. When the ratio of ships that are detained (positive class) to ships that
are not detained (negative class) is 1:99, predicting all the ships to be not detained
could achieve a 99% accuracy rate, but the model is meaningless as the detained
ships are indeed our focus.
Therefore, metrics suitable to be applied to address imbalanced problems should
be proposed.

4.2.5.1.2  Precision
‍Precision‍ aims to evaluate the proportion of predicted positive samples that are
indeed positive. It takes the following form:
TP
Precision = .
‍ TP + FP ‍
 Key elements of data-­driven models  45

For the above Example 4.2.1, ‍Precision‍ is undefined as T

‍ P + FP = 0‍ is a cautious
metric that focuses on truly positive samples.

4.2.5.1.3   Recall
‍Precision‍ is from the perspective of samples predicted to be positive. From the
perspective of samples that are actually positive, the metric ‍Recall ‍aims to evaluate
the proportion of actual positive samples that are correctly predicted. It takes the
following form:
In Example 4.2.1, Recall = 0 as TP = 0 aims to capture as many positive samples
as possible. It is also noted that if we predict all samples to be positive, we have
Recall = 1.
‍Precision‍ and ‍Recall ‍ are somewhat contradictory as they view the same prob-
lem from opposite angles: ‍Precision‍only looks at the samples predicted to be posi-
tive, while ‍Recall ‍takes all the actual positive samples into account.

4.2.5.1.4   F1 score
To reach a trade-­off between ‍Precision‍and ‍Recall ‍, the metric of F1 score, which is
the harmonic mean of ‍Precision‍and ‍Recall ‍, takes the following form:
When T ‍ P + FP = 0‍where ‍Precision‍is undefined and T ‍ P + FN = 0‍where ‍Recall ‍
is undefined, ‍F1 = 0‍. The range of F1 score is between 0 and 1, and a higher value
indicates better performance of a classification model. F1 score treats ‍Precision‍and
‍Recall ‍ equally. In case that domain knowledge of different weights for ‍Precision‍
and ‍Recall ‍should be included, a more generic ‍Fˇ ‍score can be applied, which takes
the following form:
β is chosen such that ‍Recall ‍ is considered β times as important as ‍Precision‍.
When ‍ˇ = 1,‍ ‍Fˇ ‍is degenerated to F1. When ‍ˇ > 1,‍ ‍Recall ‍has a larger impact on ‍Fˇ ;‍
when ‍ˇ < 1,‍ ‍Precision‍has a larger impact.

4.2.5.1.5   Receiver operating characteristic (ROC) curve

In many classification ML models, a value between 0 and 1 or the probability of
being positive is predicted for one sample first, and the value/probability is then
compared with a threshold (which is usually 0.5): if the value/probability is larger
than the threshold, it is predicted to be positive; otherwise, it is predicted to be nega-
tive. For samples predicted to be positive, they can be either positive (TP) or nega-
tive (FP) actually, and the rate of positive samples predicted to be positive [TP rate
(TRP)] and the rate of negative samples predicted to be positive [FP rate (FPR) are
calculated as follows:
TP FP
TPR = , FPR = .
‍ TP + FN TN + FP ‍

Given a data set (with limited samples) and a threshold, a pair of ‍(FPR, TPR)‍ can
be obtained. By changing the threshold from the smallest to the largest predicted
values/probabilities for all the samples, different ‍(FPR, TPR)‍pairs can be obtained,
and linking all the consecutive pairs together with ‍(0, 0)‍ and ‍(1, 1)‍ can produce the
46  Machine learning and data analytics for maritime studies

Figure 4.2   An illustration of ROC curves

ROC curve, where an illustration is given by the solid line in Figure 4.2. As limited
samples are contained in the test set, the solid line is ragged. The dotted diagonal
line corresponds to the ROC curve of random guessing.
Figure 4.2 shows that as the threshold decreases, more samples are predicted to
be positive, and ‍ FPR‍ and ‍ TPR‍increase simultaneously from ‍(0, 0)‍, where the thresh-
old is 1 with no sample predicted to be positive, to ‍(1, 1)‍, where the threshold is 0
with no sample predicted to be negative. A perfect classifier has ‍(FPR, TPR) = (0, 1)‍.
Therefore, the closer an ROC curve to (0,1), the better a classifier is.

4.2.5.1.6   Area under ROC curve (AUC)

When comparing two classifiers, if the ROC curve of one classifier is fully above
the other, it can be safely concluded that the performance of the former is better
than that of the latter. However, if their ROC curves intersect, it can be hard to tell
which is better. To remedy this issue, the concept of area under ROC curve, or AUC,
is proposed. It is a single value within range [‍ 0.5, 1]‍ to evaluate the performance
 Key elements of data-­driven models  47

of a classifier, and a larger value indicates a better performance. If AUC is smaller

than 0.5, one can simply swap the predicted negative and positive classes. Suppose
that the ROC curve is formulated by several pairs ‍(FPR1 , TPR1 ),‍ ‍(FPR2 , TPR2 ),‍ …,
‍(FPRn , TPRn )‍ where ‍FPRi ‍ increases from ‍(FPR1 = 0)‍ to ‍(FPRn = 1),‍ AUC can be
calculated by

4.2.5.2 Model selection
An ideal ML model should have good generalization ability, or low generalization
error, that performs well on unseen data. Therefore, there is a trade-­off between
over-­fitting and under-­fitting when an ML model is constructed: if the model is too
complex, training error can be very low or even zero. Meanwhile, the possibility
of over-­fitting the training data can be high, leading to a poor generalization abil-
ity of the ML model constructed. In contrast, a too simple ML model might lead
to under-­fitting as well as a poor generalization ability. An ideal ML model can be
selected by the metrics introduced in section 4.2.5.1 considering different scenarios
and requirements. Model selection aims to estimate the performance of different ML
models using proper metrics in order to choose the best one. To achieve effective
selection, a validation set/several validation sets that is/are mutually exclusive with
the training set and a test set should be formulated and used for model evaluation
and selection. The ML model with the lowest prediction error on the validation set(s)
(where we expect that it also has the lowest prediction error on the unseen test set)
should be selected. Common data set division methods used for model selection are
given below.
A common approach is to randomly divide the whole data set into a training set,
a validation set, and a test set, where the training set is used for model construction,
the validation set is used for model selection, and the test set should not be used until
the final ML model is constructed and the generalization error of the final chosen
model is assessed. There is no general rule to decide the number or ratio of samples
in these three sets, as it depends on the size and features of the whole data set.
Common ratios of these sets are 50:25:25, 60:20:20, and 70:15:15% To reduce the
uncertainties brought by the division of data set, multiple times of random partition
should be conducted, and the results generated should be averaged.
Another popular approach is called cross validation. After splitting a test set
from the whole data set ‍ D‍, the remaining samples in ‍ D‍, denoted by ‍ D0 ‍, are divided
into ‍k ‍ (‍k > 2‍, and the common values are 3, 5, and 10) mutually exclusive sub-
sets, i.e., ‍D0 = D1 [ D2 [ ... [ Dk ‍ and ‍Di \ Dj = ¿, i = 1, ..., k, j = 1, ..., k, i ¤ j .‍
Cross validation in this form is also called ‍k ‍-­fold cross validation. For each of the ‍k ‍
times’ training, one subset is selected as the validation set and the remaining subsets
form the training set. Then, ‍k ‍ model performance results can be obtained, and the
average performance can be calculated. An illustration of the 1‍ 0‍-­fold cross valida-
tion is given in Figure 4.3. When ‍k = n0 ‍, where ‍n0 ‍ is the size of ‍ D0 ‍, each sample is
given the opportunity to be used as the hold out validation set. This approach is also
called leave-­one-­out cross validation. In each time of training, all the samples except
the one in the validation set are used to train the ML model, and the ML models’
48  Machine learning and data analytics for maritime studies

Figure 4.3   Procedure of 1‍ 0‍-­fold cross validation

performance is validated by all samples. Therefore, the results can be regarded to

be very similar to the actual performance on the whole data set. However, when
the sample size is too large, the computational burden can be very large and even
intolerable.
The size of training set in the above approaches is more or less smaller than the origi-
nal data set. Consequently, inconsistency in data distribution due to different sample sizes
might be brought. To remedy this issue, bootstrapping that is based on bootstrap sampling
is applied for model selection. Given a data set ‍ D‍ containing ‍n‍samples, we use bootstrap
sampling to formulate a training set ‍ D N ‍by randomly selecting a sample from ‍ D‍and copy-
N ‍, and then the sample is put back to ‍ D‍, and thus this sample is also likely to be
ing it to ‍ D
picked in the next round of sampling. The above steps are repeated ‍n‍times, and finally,
the training set ‍ D N ‍also has ‍n‍samples. It is highly likely that ‍ D
N ‍has duplicate samples, and
some samples in ‍ D‍are not in ‍ D N ‍. It can be estimated that after ‍n‍rounds of sampling, the
1
possibility that one sample is not picked is (1  )n, and we have
‍ n ‍
1 1
lim (1  )n !  0.368,
‍ n!+1 n e ‍
which means that about 36.8% of the samples do not appear in ‍ D N ‍. Therefore, we
N ‍ as the validation set. The validation error
N ‍as the training set and ‍D \ D
can use ‍ D
obtained in this way is also called out-­of-­bag error.

4.2.6   Model refinement

We mainly introduce hyperparameter tuning in this subsection for refining the per-
formance of ML models. A hyperparameter is a parameter used to control the over-
all learning process and is manually set before model learning. Hyperparameters
are important to ML models as model performance is highly dependent on them.
Hyperparameters from an ML model itself are responsible to control the capacity of
 Key elements of data-­driven models  49

the model, that is, how flexible the model is in fitting the data. Therefore, they can
be used to control over-­fitting and to avoid under-­fitting. Although default values of
the hyperparameters are given in ML models implemented by open source packages,
finding their best values given a specific data set is not a trivial task, especially in
ML models that have multiple hyperparameters that interact in nonlinear ways.
Hyperparameter tuning aims to search for the best values for a set of hyperpa-
rameters that result in the best model performance on a given data set. Denote the
number of hyperparameters that need to be optimized by ‍p.‍ The tuning procedure
starts by defining a search space for the ‍p‍ hyperparameters to be tuned, and then
different hyperparameter settings constructed from the search space are evalu-
ated by a tuning algorithm. Finally, the hyperparameter setting leading to the best
model performance on the validation sets is output, which is used to train the final
model.
The most commonly used hyperparameter tuning algorithm is grid search, which
picks out a grid of values of the hyperparameters concerned and evaluates them one
by one. It is also noted that many hyperparameters take values in the range of real
numbers. However, in grid search, a range of their values and the searching interval
are first given, and thus finite candidate values are contained in the search space.
For example, for an ML model with ‍p‍hyperparameters h1 to hp, the search space is
n o
v
h1 2 h11 , h21 , ..., h11 ,
..
.
n vp
o
1 2
hp 2 hp , hp , ..., hp .
‍ ‍
Then, a grid of ‍v1  ...  vp‍ hyperparameter settings can be formulated with each
constituted by one possible value of h1 to hp. All hyperparameter settings are then
tested, and the hyperparameter setting with the best performance on the validation
set or in cross validation is returned.
Grid search is quite simple and intuitive. However, it is very expensive in terms
of computation power. When the ML model is complex and the number of hyper-
parameters is large, the computation time might become unaffordable. Instead of
searching for the entire search space, random search is proposed which only eval-
uates a certain number of random hyperparameter settings (the number is much
smaller than ‍v1  ...  vp)‍ in the grid. Theoretically, the performance of the best
hyperparameter setting found by grid search is highly likely to be worse than that
found by grid search. Nevertheless, in many cases of public datasets, Bergstra and
Bengio (2012) [5] found that random search can perform about as well as grid search
with much less computation power needed.
In recent years, smart hyperparameter tuning algorithms are developed. Unlike
the grid search and random search algorithms that form the candidate hyperparame-
ter settings in advance and then evaluate all or part of them one by one, smart hyper-
parameter tuning algorithms only pick a few hyperparameter settings and evaluate
their quality, and then decide where to sample next. The drawback is that the former
two tuning algorithms can be parallelizable, while the latter is sequential. Popular
50  Machine learning and data analytics for maritime studies

algorithms are Bayesian optimization [6], derivative-­free optimization [7], and ran-
dom forest smart tuning [8].

4.2.7 Model assessment, interpretation/explanation, and conclusion

After selecting the most suitable model whose performance is refined by hyper-
parameter tuning, the final chosen model’s performance should be assessed by a
hold-­out test set that should never be seen and used in the above processes (e.g.,
in model construction, selection, and refinement). The model performance, includ-
ing its working mechanism and prediction results, might need to be further inter-
preted and explained. ML model interpretation and explanation is another large and
increasingly important topic, and we will not cover it in this chapter. Readers are
referred to Reference 9. Finally, based on the prediction model constructed and the
prediction results, conclusions are generated and managerial insights to improve the
efficiency of operation and business are derived.

References

[1] Tu E., Zhang G., Rachmawati L., Rajabally E., Huang G.B. ‘Exploiting AIS data
for intelligent maritime navigation: a comprehensive survey from data to meth-
odology’. IEEE Transactions on Intelligent Transportation Systems. 2017;19
(5):1559–82.
[2] Yang D., Wu L., Wang S., Jia H., Li K.X. ‘How big data enriches maritime
research—a critical review of automatic identification system (AIS) data ap-
plications’. Transport Reviews. 2019;39(6):755–73.
[3] Yan R., Wang S., Peng C. ‘An artificial intelligence model considering data
imbalance for SHIP selection in port state control based on detention prob-
abilities’. Journal of Computational Science. 2021;48:101257.
[4] Solorio-­Fernández S., Carrasco-­Ochoa J.A., Martínez-­Trinidad J.F. ‘A review
of unsupervised feature selection methods’. Artificial Intelligence Review.
2020;53(2):907–48.
[5] Bergstra J., Bengio Y. ‘Random search for hyper-­parameter optimization’.
Journal of Machine Learning Research. 2012;13(2):281–305.
[6] Snoek J., Larochelle H., Adams R.P. ‘Practical bayesian optimization of
machine learning algorithms’. Advances in Neural Information Processing
Systems. 2012;25.
[7] Conn A.R., Scheinberg K., Vicente L.N. Introduction to derivative-­free opti-
mization [online]. SIAM; 2009 Jan. Available from http://epubs.siam.org/​doi/
book/10.1137/1.9780898718768
[8] Hutter F., Hoos H.H., Leyton-­Brown K. ‘Sequential model-­based optimization
for general algorithm configuration’. International Conference on Learning
and Intelligent Optimization; 2011. pp. 507–23.
[9] Interpretable machine mearning: a guide for making black box models ex-
plainable [online]. 2022 Mar 29. Available from https://christophm.github.​io/
interpretable-ml-book [Accessed 31 Mar 2022].
 Chapter 5
Linear regression models

Linear regression aims to learn a linear model that can predict the target using the
features as accurately as possible. The assumption of linear regression models is
that the target is linearly correlated with the features, i.e., the regression function
‍E(y|x)‍is linear in ‍x‍, where ‍E()‍denotes expectation. If the assumption can (almost)
be satisfied, linear regression can be comparable or can even outperform fancier
non-­linear models. The linear regression model is one of the most classic models for
prediction tasks, and it is still widely used in the computer and big data era, thanks
to its intuitiveness and interpretability in particular. In the following sections, we
first introduce simple linear regression models (with a single feature) and the least
squares method, which aims to find the optimal parameters of a linear regression
model by minimizing the sum of squares of the residuals. Then, we discuss multiple
linear regression (with more than one feature) and its extension. Finally, we intro-
duce shrinkage linear regression models.

5.1 Simple linear regression and the least squares

Simple linear regression uses only one feature to predict the target. For example,
we use ship age to predict the number of deficiencies of a PSC inspection. Denote
the training set with ‍n‍ samples by ‍D = f(x1 , y1 ), (x2 , y2 ), ..., (xn , yn )g‍ and the feature
vector by ‍x ‍. Simple linear regression aims to develop a model taking the following
form:
‍ yO i = wxi + b ,‍
where ‍yO i ‍is the predicted target for sample i‍ ‍, ‍w ‍is the parameter weight and ‍b‍is the
bias. ‍w ‍and ‍b‍need to be learned from ‍ D‍. Then, a natural question is: what are good
‍w ‍and ‍b‍? Or in other words, how to find the values of ‍w ‍and ‍b‍such that the predicted
target is as accurate as possible? The key point of developing a simple linear regres-
sion model is to evaluate the difference between ‍yO i ‍and yi, i‍ = 1, ..., n‍using the loss
function and to adopt the values of ‍w ‍ and ‍b‍ that minimize the loss function. In a
regression problem, the most P commonly used loss function is the mean squared error
(MSE), where ‍MSE = 1n ni=1 (yi  yOi )2‍. Therefore, the learning objective of simple
linear regression is to find the optimal ‍(w , b )‍ such that the MSE is minimized.
The above idea can be presented by the following mathematical functions:
52  Machine learning and data analytics for maritime studies
P
n
(w , x ) = arg min (yi  yOi )2
(w,b) i=1
Pn . (5.1)
= arg min (yi  wxi  b)2
‍ (w,b) i=1 ‍
This idea is called the least squares method. The intuition behind it is to minimize
the sum of lengths of the vertical lines between all the samples and the regression
line determined by ‍w ‍and ‍b‍. It can easily be shown that ‍MSE ‍is convex in ‍w ‍and ‍b‍,
and thus ‍(w , b )‍can be found by
@MSE P
n  
= 2( xi wxi  (yi  b) ) = 0
@w Pi=1
n Pn . (5.2)
i=1 yi (xi  n xi )
1

) w = Pn 2 1 Pn i=1 2


‍ i=1 xi  n ( i=1 xi ) ‍
The optimal ‍w‍is first found by Equation (5.2), and then it can be used to calculate
the optimal value of ‍b‍, denoted by ‍b‍, as follows:
@MSE Pn
= 2( w xi + b  yi ) = 0
@b i=1
1X
n . (5.3)
)b =
(yi  w xi )
‍ n i=1 ‍
Simple linear regression can easily be realized by scikit-­learn API [1] in Python.
Here is an example of using ship age to predict ship deficiency number using simple
linear regression.
Example 5.1: In this problem, the only feature is ship age and the target is
ship deficiency number. ‍(w , b )‍ is estimated on the training set using the
LinearRegression method in scikit-­learn API. The core code is as follows:

‍ ‍

Therefore, the relationship between ship age and deficiency number learned by
simple linear regression can be presented in:
‍ deficiency number = 0.2648  age + 1.‍ (5.4)
Model performance is then validated on the hold-­out test set, which can also
easily be achieved by the scikit-­learn API. Finally, we have ‍MSE = 20.3471‍,
‍RMSE = 4.5108‍, ‍MAE = 2.9650‍, and ‍R2 = 0.0713‍. Recall that the definitions of
MSE, RMSE, and MAE are given in section 4.2.4 of chapter 4. Note that in
this example, ‍ R2 ‍ is very small, indicating that the model is highly likely to be
ineffective, and simple linear regression might not be suitable to address this
problem.
 Linear regression models  53

5.2 Multiple linear regression

In more general cases, the feature dimension denoted by ‍m‍ is more than one, i.e.,
two or more features are used to predict the target. For example, we use a linear
regression model with five features: ship age, type, GT, last inspection time, and last
deficiency number to predict the ship deficiency number of the current PSC inspec-
tion, and this model is called multiple linear regression. Mathematically, multiple
linear regression can be written as follows:
yO i = w1 xi1 + w2 xi2 + ... + wm xim + b
= xi w + b,
(5.5)
xi = (xi1 , xi2 , ..., xim ),
‍ w = (w1 ; w2 ; ...; wm ). ‍
To unify the parameters, Equation (5.5) can be written as
yO i = xQ i w,
Q
xQ i = (xi1 , xi2 , ..., xim , 1), (5.6)
‍ wQ = (w1 ; w2 ; ...; wm ; b).‍
In a training set ‍ D‍ with ‍n‍samples, the matrix form of Equation (5.6) is
0 1
0 1 0 1 w1
yO 1 x11 x12    x1m 1 B C
B C B CB w C
ByO 2 C Bx21 x22    x2m 1CB 2 C
B C B CB .. C
B . C=B . .. .. .. .. CB . C. (5.7)
B .. C B .. . . . .C B C
@ A @ AB C
@wm A
yO n xn1 xn2    xnm 1
‍ b ‍

or ‍yO = xQ wQ ‍. The dimension of ‍yO ‍ is ‍n  1‍, the dimension of ‍xQ ‍ is ‍n  (m + 1)‍, and
the dimension of ‍wQ ‍ is ‍(m + 1)  1‍. A closed-­form solution of the optimal ‍wQ ‍ can
be found by the least squares method but the case is more complex than that in
simple linear regression. The MSE is calculated by ‍MSE = (y  xQ w) Q ‍, and
Q T (y  xQ w)
its derivative to ‍wQ ‍is
@MSE
= 2QxT (Qxw
Q  y). (5.8)
‍ @wQ ‍
If the matrix ‍xQ T xQ ‍is a full-­rank matrix or a positive definite matrix, ‍wQ ‍can be calcu-
lated by setting Equation (5.8) to zero, and we can have

‍ wQ  = (QxT xQ )1 xQ T y.‍ (5.9)

here ‍(QxT xQ )1‍ is the inverse matrix of ‍xQ T xQ ‍. The optimal multiple linear regression
model solved by least squares is

‍ yO i = xQ i (QxT xQ )1 xQ T y.‍ (5.10)

54  Machine learning and data analytics for maritime studies

If ‍xQ T xQ ‍ is not a full-­rank matrix, which indicates that the features outnumber the
samples, then there will be several ‍wQ ‍, and the optimal one can be selected by intro-
ducing regularization in section 5.4.
Multiple linear regression can also be realized by scikit-­learn API [1] in Python.
Here is an example of using ship age, type, GT, last inspection time, and last defi-
ciency number, whose details are provided in section 2.3 of chapter 2, to predict ship
deficiency number using multiple linear regression.
Example 5.2: Ship type is a nominal categorical feature. We first use one-­hot
encoding to encode this feature to six features: type_bulk_carrier, type_container_
ship, type_general_cargo/multipurpose, type_passenger_ship, type_tanker, and
type_other. For ships without the last inspection record, we set ‍1‍to their features
“last_inspection_time” and “last_deficiency_no.” Then, we have a total of ten fea-
tures, which are of varied ranges. For example, in the whole dataset, age ranges from
0 to 46, while GT ranges from 0 to 210,678. Therefore, min-­max scaling (readers are
referred to section 4.2.4 of chapter 4) is used for feature scaling. It should be noted
the min-­max scaler should be first fit on the features in the training set, and then
applied to the features in the training and test sets. Then, the training set and test set
after feature scaling using a min-­max scaler are used to train and test the multiple
linear regression model. Similar to Example 5.1, ‍wQ ‍can be estimated on the training
set using the LinearRegression in scikit-­learn API. In this example, the ship defi-
ciency number is predicted by Equation (5.11), where the features have been scaled
by the min-­max scaler:

deficiency number = 8.1786  age + 0.1671  type_bulk_carrier
 0.4938  type_container_ship + 3.9357  type_general_cargo/multipurpose
+ 0.9053  type_other  2.9534  type_passenger_ship
 1.5609  type_tanker  2.9414  GT
‍ 2.8232  last_inspection_time + 13.6901  lastdeficiency_no + 2.3880. ‍
(5.11)

The weights of the features show their contributions to the number of ship deficien-
cies detected. For example, ship deficiency number increases rapidly as ship age or
the last deficiency number increases. In contrast, given all other conditions equal,
a larger ship has less deficiencies. The longer the last inspection time, the fewer
deficiencies detected. This is because ships in worse conditions are more likely to be
selected for inspection in the current SRP ship selection scheme, and thus they have
shorter inspection intervals. For different types of ships, being types general_cargo/
multipurpose, other, and bulk carrier increases the deficiency number, while being
passenger ship, tanker, and container ship decreases the deficiency number. The
above analysis shows that linear regression is totally explainable and its compliance
with domain knowledge can easily be validated.
 Linear regression models  55

Model performance is then validated on the hold-­out test set, and we have
‍MSE = 15.9987‍, ‍RMSE = 3.9998‍, ‍MAE = 2.5948‍, and ‍R2 = 0.2698‍. It is obvious
that the performance of the multiple linear regression model for ship deficiency
number prediction is much better than that of the simple linear regression model
in Example 5.1, as many more features that can influence the target are considered.

5.3 Extensions of multiple linear regression

This section introduces two popular extensions of multiple linear regression: pol-
ynomial regression and logistic regression. As a relatively simple and basic ML
model considering a linear relationship between the features and the target, multiple
linear regression is highly likely to lead to the problem of underfitting. To address
this problem, one viable approach is to incorporate non-­linearity into the model. To
retain the linear function form (and thus model interpretability), higher order terms
of the original features (e.g. quadratic, cubed, and product terms) are formed and
included in the multiple linear regression model. The model formulated is called
polynomial regression. Besides, multiple linear regression can only address regres-
sion problems. To extend multiple linear regression to address classification prob-
lems, the continuous output given by multiple linear regression is mapped to classes
using a surrogate function. For binary classification problem, logistic function is a
commonly used mapping function, and the classification model is called logistic
regression. Details of the two extensions of multiple linear regressions are given in
the following subsections.

5.3.1 Polynomial regression
As linear regression assumes a linear relationship between the original features and
the target by constructing a straight line, underfitting occurs if the features and the
target take a non-­linear relationship, where the complexity of the model needs to
be improved to enhance model performance. One approach is first to add powers
to the original features to form new (and more complex) features, and then to form
a linear function between the new features and the target. This approach is called
polynomial regression, and the features constructed are called polynomial features.
For example, given two features ‍(X1 , X2 )‍and power degree 2, each of them will be
transformed to 3 features with orders from 0 to 2 (note that both features have the
same value 1 when the order is 0), respectively, in addition to their product term.
Therefore, the new features will be ‍(1, X1 , X21 , X2 , X22 , X1 X2 ).‍ Applying polynomial
regression to Example 5.2 yields Example 5.3.
Example 5.3: The 6 binary features encoded from the nominal feature ship type
are not processed and are directly used in polynomial regression. Min-­max scaling
is applied to numerical features age, GT, last inspection time, and last deficiency
number, and their polynomial features of degree 2 are constructed by scikit-­learn
API using the following codes:
56  Machine learning and data analytics for maritime studies

‍ ‍

Then, a total of 15 new features can be constructed from the 4 original features,
including 1 constant term, 4 linear terms, 4 quadratic terms, and 6 product terms.
Totally, there are 21 features (15 numerical features and 6 categorical features).
Then, multiple linear regression is applied to all the 21 features. We omit the specific
form as it is similar to that of Example 5.2. Model performance is then validated on
the hold-­out test set, and we have ‍MSE = 15.4055‍, ‍RMSE = 3.9250‍, ‍MAE = 2.5140‍,
and ‍R2 = 0.2969‍. The performance of polynomial regression is better than that of
the multiple linear regression model in Example 5.2 and much better than that of the
simple linear regression model in Example 5.1, as a more complex feature form is
used.

5.3.2 Logistic regression
If the prediction target is whether a ship is detained in an inspection, which is a
binary variable with “1” indicating ship detention and “0,” otherwise. Neither
simple nor multiple linear regression model can be directly applied to this clas-
sification problem, as the output is continuous and unbounded, while we expect
the output to be categorical and bounded. An intuitive method is to set a threshold
to predict the probability of ‍y = 1‍ given the input features ‍ x‍, i.e., ‍P(y = 1|x).‍ The
unit-­step function is a popular method to map a continuous output (denoted by ‍ z ‍)
to a probability (denoted by ‍e y ),
‍ which takes the following form as shown in
Figure 5.1.
However, Figure 5.1 shows that the final output given by the unit-­step function
is discontinuous, making it hard to be optimized. Therefore, a continuous, mono-
tonic, and differentiable surrogate function of the unit-­step function called logistic
function taking the following form is used:
1
e
y= , (5.12)
‍ 1 + ez ‍

here ‍z = xQ wQ ‍is the continuous output given by a multiple linear regression model. An
illustration of the logistic function is shown in Figure 5.2.
Equation (5.12) can also be transformed as follows:
 Linear regression models  57

Figure 5.1   An illustration of unit-­step function

1
e
y =
1 + ez (5.13)
e
y
)z = xQ w
Q = ln .
‍ 1 ey ‍

In Equation (5.13), ‍e y ‍ is the probability of a sample with features ‍x‍to be of class

e
y
“1” and 1‍ e y ‍ is the probability to be of class “0.” Therefore, is the relative
‍1 e y‍
e
y
probability of sample ‍x‍to be of class “1,” which is called odds. ln is the natu-
‍ 1 e y‍
ral log of odds, and is called log odds, or logit. Therefore, Equation (5.13) can be
interpreted as using the output of a multiple linear regression model to approximate
the log odds, so as to map a continuous target to a probability.
Similar to simple linear regression and multiple linear regression, the optimal
values of ‍wQ ‍need to be estimated by minimizing the estimation error. Equation (5.13)
can be further written as
58  Machine learning and data analytics for maritime studies

Figure 5.2   An illustration of the logistic function

e
y
z = xQ wQ = ln
1 ey
P(y = 1|x) (5.14)
) xQ wQ = ln .
‍ P(y = 0|x) ‍
and we can have
exQwQ
P(y = 1|x) = xQwQ ,
1+e (5.15)
1
P(y = 0|x) = xQwQ .
‍ 1 + e ‍
Then, the maximum likelihood method [2] can be used to estimate ‍wQ ‍ to maximize
the probability that the predicted target of each example is equal to its actual tar-
get. Algorithms in numerical optimization, such as the gradient descent method and
Newton method can be used to obtain the optimal ‍wQ ‍, i.e., ‍wQ ‍. We use the same fea-
tures of Example 5.2 to predict ship detention. The procedure and results are shown
in Example 5.4.
 Linear regression models  59

Example 5.4: Min-­max scaling is also first applied to numerical features age,
GT, last inspection time, and last deficiency number. However, one critical issue
in this example is that ship detention is a rare event. For example, among all the
1,592 samples in the training set, the target of only 85 of them is “1,” indicating that
the detention rate is about 5.34%. To overcome the problem of data imbalance, we
first use the synthetic minority oversampling technique (SMOTE) to oversample the
minority class, with the aim to construct a balanced dataset. Then, logistic regression
is applied to predict ship detention. We use a pipeline combining SMOTE and logis-
tics regression to first construct a balanced data set, and then predict ship detention
based on scikit-­learn API with main code as follows:

‍ ‍

Then, the constructed model is validated on the hold-­out test set containing 399
examples (15 examples with detention), and the following confusion matrix given in
Table 5.1 is obtained. The average ‍Precision‍, ‍Recall ‍, and ‍F1‍scores are 0.55, 0.79,
and 0.54, respectively.

5.4 Shrinkage linear regression models

Shrinkage in linear regression models is often achieved by imposing a penalty on

coefficients to reduce the absolute values of coefficients or the number of valid fea-
tures in the regression model, so as to reduce the problem of overfitting, enhance the
prediction accuracy, and increase model interpretability (as less features are finally
contained in the model, meaning that less important features can be excluded from
the model and the more important ones are left). In this section, we introduce two
popular shrinkage linear regression models: ridge regression and LASSO regression.

Table 5.1  Confusion matrix for ship detention prediction using logistic
regression

Predicted value/Actual value Detained Not detained

Detained 12 88
Not detained 3 296
60  Machine learning and data analytics for maritime studies

5.4.1 Ridge regression
Ridge regression imposes a penalty on the size of the regression coefficients using
L2 regularization, where the loss function takes the following form:
Pn Pm P m
l = (yi  b  xij wj )2 +  wj2 , where  > 0. (5.16)
‍ i=1 j=1 j=1 ‍
‍‍is a complexity parameter to control the degree of shrinkage: a larger ‍ ‍means a
greater amount of shrinkage. The objective of ridge regression is to find the optimal
‍w‍such that
( )
Pn Pm Pm
w = arg minw

(yi  b  xij wj ) +  wj .
2 2
(5.17)
‍ i=1 j=1 j=1
‍
This is equivalent to solving the following optimization problem:
( )
Pn Pm
w = arg minw

(yi  b  xij wj ) ,
2

i=1 j=1
(5.18)
Pm
s.t. wj  t,
2

‍ j=1 ‍
here there is a one-­to-­one relationship between ‍ ‍ and ‍t ‍, and the size constraint
on the parameters (i.e. constraint on parameter values) is imposed explicitly in
Equation (5.18). Ridge regression is effective to alleviate the problem of high vari-
ance brought about by correlated variables in multiple linear regression by shrinking
coefficients close to (but not exactly) zero. It is also noted that bias ‍b‍, which is not
directly related to the parameters, is excluded from the penalty terms, as they aim
to regularize the coefficients of parameters. Its value should also be determined in
Equation (5.18). An example of using ridge regression to predict ship deficiency
number using the features of Example 5.2 based on scikit-­learn API is as follows.
Example 5.5: Min-­max scaling is also first applied to numerical features age,
GT, last inspection time, and last deficiency number. Ridge regression with hyper-
parameter tuning for ‍ ‍based on 5-­fold cross-­validation can easily be implemented
by the RidgeCV method provided by scikit-­learn API. The main code is as follows:
‍

‍
 Linear regression models  61

The constructed ridge regression model for ship deficiency number prediction
takes the following form:
deficiencynumber = 8.1786  age + 0.1671  type_bulk_carrier
 0.4938  type_container_ship + 3.9357  type_generalc argo/multipurpose
+ 0.9053  type_other  2.9534  type_passenger_ship
 1.5609  type_tanker  2.9414  GT
‍  2.8232  last_inspection_time + 13.6901  lastdeficiency_no + 2.3880. ‍
(5.19)
Compared to Equation(5.11) for ship deficiency number prediction using multiple
linear regression, it can be seen that the absolute weight values associated with the
parameters are smaller in Equation(5.19) in ridge regression. Meanwhile, the relative
importance of the parameters is similar in both models. Model performance is then
validated on the hold-­out test set, and we have ‍MSE = 15.9981‍, ‍RMSE = 3.9998‍,
‍MAE = 2.5948‍, and ‍R2 = 0.2698‍, which is slightly better than the performance of
the multiple linear regression model as shown in Example 5.2.

5.4.2 LASSO regression
LASSO regression imposes L1 regularization on the size of the regression coef-
ficients, whose loss function takes a similar form of ridge regression except for the
regularization term:
Pn Pm Pm
l= (yi  b  xij wj )2 +  |wj |, where  > 0. (5.20)
‍ i=1 j=1 j=1 ‍
And it is equivalent to solving the following optimization model to find the optimal ‍w‍:
( )
Pn Pm
w = arg minw

(yi  b  xij wj ) ,
2

i=1 j=1
(5.21)
Pm
s.t. |wj |  t.
‍ j=1 ‍

The optimal value of ‍b‍can also be determined by Equation (5.21). As the L1 regu-
larization term with absolute operation involved is used in LASSO regression, it can
lead to zero coefficient, which means that some features are totally ignored by the
regression model. The larger value of ‍ ‍, the fewer features are retained. Therefore,
LASSO regression can also be used for feature selection in addition to overfitting
control. An example of using LASSO regression to predict ship deficiency number
using the features of Example 5.2 based on scikit-­learn API is as follows.
Example 5.6: Min-­max scaling is also first applied to numerical features age,
GT, last inspection time, and last deficiency number. Like ridge regression, LASSO
regression with hyperparameter tuning for ‍ ‍ based on 5-­fold cross validation can
easily be implemented by the LassoCV method provided by scikit-­learn API. The
main code is as follows:
62  Machine learning and data analytics for maritime studies

‍ ‍

The constructed LASSO regression model for ship deficiency number predic-
tion takes the following form:
deficiency number = 8.0205  age + 0  type_bulk_carrier
 0.6216  type_container_ship + 3.7846  type_general_cargo/multipurpose
+ 0.7083  type_other  2.7820  type_passenger_ship
 1.6691  type_tanker  2.9266  GT
‍  1.9954  last_inspection_time + 13.6331  last_deficiency_no + 2.5125. ‍ (5.22)

Equation (5.22) shows that the coefficient of feature type_bulk_carrier is shrunk

to exactly 0, indicating that the LASSO algorithm can be used for feature selec-
tion. However, it is noted that the model performance of LASSO is slightly
worse than multiple linear regression and ridge regression, with ‍MSE = 16.0133‍,
‍RMSE = 4.0017‍, ‍MAE = 2.5908‍, and ‍R2 = 0.2691‍. This shows that LASSO may not
be a suitable method for ship deficiency number prediction, as in this problem, the
number of examples is much larger than the number of features and these features
are not highly correlated with each other.
From the results of Examples 5.5 and 5.6, it can be seen that ridge regression
tends to produce small but non-­zero weights, while LASSO tends to produce sparse
weights (to exactly zero). Therefore, when there are too many features, especially
when the number of features is larger than the number of examples, LASSO regres-
sion might be more suitable to be applied for feature selection. Otherwise, ridge
regression might be more suitable to be applied for feature coefficient regulation.

References

[1] Pedregosa F., Varoquaux G., Gramfort A, et al. ‘Scikit-­learn: machine learning
in python’. Journal of Machine Learning Research. 2011;12:2825–30.
[2] Rossi R.J. Mathematical statistics [online]. Hoboken, NJ: John Wiley & Sons;
2018 Jul 18. Available from http://doi.wiley.com/10.1002/9781118771075
 Chapter 6
Bayesian networks

This chapter introduces the basics of Bayesian network (BN) classifiers that are used
to address classification problems. Naive Bayes classifier is first presented, where a
simplified (but unrealistic) assumption that the features are conditionally independ-
ent and are of equal importance is made. To weaken the assumption so as to improve
the classification accuracy, semi-­naive Bayes classifiers are then presented, where
part of the dependencies between the features is considered. Finally, BN in more
general form is introduced.

6.1  Naive Bayes classifier

Consider a classification problem where the target has a total of ‍ K ‍ possible

values denoted by ‍C = fc1 , ..., cK g‍. We have a set of samples (observations)
m
‍D = f(xi , yi ), i = 1, ..., ng‍, where ‍xi 2 R ‍ and ‍yi 2 C‍. For observations whose set
of features is ‍xi ‍, due to the limited types of features considered and the errors in
measurement, they may be in different classes in the dataset ‍ D ‍. Therefore, for an
unseen observation with a feature set ‍xi ‍, the probability of this observation to be
of class ck can be estimated by Bayes’ rule as follows:
P(xi , ck )
P(ck |xi ) =
P(xi )
P(ck )P(xi |ck ) (6.1)
= ,
‍ P(xi ) ‍
where ‍P(ck )‍ is the prior probability of class ck, which can be calculated directly
from the targets of the given dataset. ‍P(ck )‍ is the weight given to the class con-
ditional probability or the likelihood of sample ‍xi ‍ given class ck, i.e., ‍P(xi |ck )‍ is
the evidence, which is the occurrence probability of the observation ‍xi ‍ and acts
as a normalizing constant for the joint probability ‍P(xi , ck )‍calculated. The main
reason of using is that the dataset is limited. Consequently, for a given example
‍xi ‍, there can be rare or even no example with the same feature values as ‍xi ‍in the
training set especially when there are a lot of features. In naive Bayes classifier,
a strong assumption is made to simplify the prediction of ‍P(ck |xi )‍: for a known
class, all the features are assumed to be independent and of equal importance.
Therefore, (6.1) can be written as
64  Machine learning and data analytics for maritime studies

P(xi , ck )
P(ck |xi )
=
P(xi )
P(ck ) Y . (6.2)
m

= P(xij |ck )
P(xi ) j=1
‍ ‍
The predicted target given the set of features ‍xi ‍ should be the class leading to
the largest value of Equation (6.2). For a given ‍xi ‍, ‍P(xi )‍ is fixed. The classifica-
tion problem is turned to estimate the product of ‍P(ck )‍and the class conditional
probability ‍P(xi |ck )‍, and find the class ck leading to the largest product as the
predicted target ‍yO i ‍. Mathematically, this process can be presented as
Q
m
yO i = arg maxck 2C P(ck )P(xij |ck ) . (6.3)
‍ j=1 ‍
|Dck |
‍P(ck )‍ can be calculated by P(ck ) = |D| , where |D ‍ ck |‍ is the number of samples
‍ ‍
in class ck, |D|
‍ ‍ is the total number of samples in the dataset, and ‍P(xij |ck )‍ is the
conditional probability of each feature ‍j ‍, ‍j = 1, ..., m‍ of sample ‍xi ‍. Especially,
for categorical features, we denote the number of samples belonging to class ck
and taking value ‍j0 ‍for feature ‍j ‍by ‍Dck ,j,j0 .‍ Then, ‍P(xij |ck )‍can be calculated by

Dck ,j,j0
P(xij |ck ) = . (6.4)
‍ |Dck | ‍

We use an example considering five features, namely ship age (age), type, last
inspection time (l_ins_time), last deficiency number (l_def_no), and last inspec-
tion state (whether a ship is detained in the last inspection; l_ins_state) to predict
whether a new coming ship has six or more deficiencies. Therefore, as the tar-
get variable, the deficiency number (denoted by def_no for short) is decoded to
two states: 0to5 and 6+. The training set contains 200 random samples from the
whole port state control (PSC) dataset.
Example 6.1 Age, l_ins_time, and l_def_no are numerical features. For the
sake of simplicity, we treat them as categorical features by discretizing their
values. The values for age after discretization are 0to5, 6to10, 11to15, and 16+.
The values for l_ins_time after discretization are no_inspection, 0to5, 6to10,
11to15, and 16+. The values for l_def_no after discretization are no_inspec-
tion, 0to5, and 6+. For the two categorical features type and l_ins_state, they
have five states (bulk_carrier, container_ship, general_cargo/multipurpose,
tanker, and other) and three states (no_inspection, no_detention, and detention),
respectively. For the three features related to last inspection, namely l_ins_time,
l_def_no, and l_ins_state, if a ship has no last inspection record in the database,
its state is set to “no inspection.” According to Equation (6.3), we first calculate
the prior probability ‍P(ck )‍for each class ‍ck 2 C ‍. The frequency and distribution
of the target are shown in Table 6.1.
Then, the conditional probability table, or the CPT, for each feature given the
target is calculated and presented as follows in Tables 6.2–6.6.
 Bayesian networks  65

Table 6.1   Frequency and distribution of def_no

State of def_no Frequency Proportion (%)

0to5 154 76.733
6+ 46 23.267

Suppose here comes a 3-­year-­old (the value of age is 0to5) tanker ship (the
value of type is tanker), the last inspection time is four months ago (the value of
l_ins_time is 0to5) with eight deficiencies identified (the state of l_def_no is 6+)
and is not detained (the state of l_ins_state is no detention). Then, the product of
the probability that the ship has 6+ deficiencies and the conditional probabilities of
the feature values given the condition that deficiency_no is 6+ can be calculated by

P(def_no = 6+)  P(type = tanker|def_no = 6+)  P(age = 0to5|def_no = 6+)

P(l_ins_time = 0to5|def_no = 6+)  P(l_def_no = 6 + |def_no = 6+)
P(l_ins_state = nodetention|def_no = 6+)
= 0.23267  0.09804  0.08  0.4902  0.26531  0.87755
 ‍  2.0827  104 . (6.5)‍

Similarly, the product of the probability that the ship has 0 to 5 deficiencies and the
conditional probabilities of the feature values given the condition that the state of
deficiency_no is 0to5 can be calculated by

P(def_no = 0to5)  P(type = tanker|def_no = 0to5)  P(age = 0to5|def_no = 0to5)

P(l_ins_time = 0to5|def_no = 0to5)  P(l_def_no = 6 + |def_no = 0to5)
P(l_ins_state = nodetention|def_no = 0to5)
= 0.76733  0.11950  0.25949  0.42767  0.05733  0.94268
 ‍  5.4995  104 . (6.6)
‍

Table 6.2   Frequency and distribution of age

State of age
State of
def_no 0to5 6to10 11to15 16+
0to5 40 (25.949%)* 57 (36.709%) 32 (20.886%) 25 (16.456%)
6+ 3 (8%) 12 (26%) 9 (20%) 22 (46%)
*The format of this cell is “frequency (proportion).” This applies for the remaining cells of this table and
the following tables.
66  Machine learning and data analytics for maritime studies

Table 6.3   Frequency and distribution of l_def_no

Value of l_def_no

State of def_no 0to5 6+ No inspection

0to5 141 (90.446%) 8 (5.733%) 5 (3.822%)
6+ 32 (67.347%) 12 (26.531%) 2 (6.122%)

Table 6.4   Frequency and distribution of l_ins_state

Value of l_ins_state

State of def_no Detention No detention No inspection

0to5 147 (94.268%) 2 (1.911%) 5 (3.822%)
6+ 42 (87.755%) 2 (6.122%) 2 (6.122%)

Table 6.5   Frequency and distribution of l_ins_time

Value of l_ins_time
State of
def_no 0to5 6to10 11to15 16+ No inspection
0to5 67 (42.767%) 55 (35.220%) 17 (11.321%) 10 (6.918%) 5 (3.774%)
6+ 24 (49.020%) 12 (25.490%) 2 (5.882%) 6 (13.725%) 2 (5.882%)

As ‍5.4995  104‍is larger than ‍2.0827  104‍, it is predicted that the coming tanker
ship has 0 to 5 deficiencies.
The above classification process can be graphically shown in the form of
a BN, which is a directed acyclic graph consisting of a set of nodes and a set
of directed arcs. Nodes in a BN can be features, latent variables (i.e. variables
that are not directly observed), and the target to be predicted. The target is the

Table 6.6   Frequency and distribution of type

Value of type

State of general_cargo/
def_no Bulk carrier Container ship multipurpose Tanker Other
0to5 12 (8.176%) 110 (69.811%) 8 (5.660%) 18 (11.950%) 6 (4.403%)
6+ 12 (25.490%) 13 (27.451%) 11 (23.529%) 4 (9.804%) 6 (13.726%)
 Bayesian networks  67

Figure 6.1  Construction of a naive Bayes classifier for ship deficiency condition
prediction

Figure 6.2 The constructed naive Bayes classifier for ship deficiency condition
prediction

class variable, and the features are the attribute variables. The directed arcs con-
necting two nodes represent conditional dependencies, where the node at the
tail of an arc is the parent node and the node at the head is the child node.
A parent node represents the condition, and its child node is the consequence of
that condition. The BN model to address the ship deficiency number prediction
problem in Example 6.1 can be constructed by the software Netica* and is shown
in Figure 6.1. A box represents a node, and the title in the first line of a box is the
name of a feature or the target. The following lines are their values (or states).
The marginal distributions of the features or the target are given by the belief
bars on the right of the states.
For the coming tanker ship, clicking each of its feature values in Figure 6.1 yields
the probability of the ship to have 0 to 5 deficiencies or 6 and more deficiencies as

* 
https://www.norsys.com/netica.html
68  Machine learning and data analytics for maritime studies

shown by the states of node “deficiency_no” in Figure 6.2. The probability of this

ship to have 0 to 5 deficiencies is 72.5%, and the probability of this ship to have 6 or
more deficiencies is 27.5%. Therefore, it is predicted that the ship has no more than
five deficiencies in the current inspection.

6.2  Semi-naive Bayes classifiers

As estimating ‍P(Ex|ck )‍, ‍k = 1, ..., K ‍ is non-­trivial, a strong assumption that the

features are independent given the class variable is made in the naive Bayes
classifier. Although it is an oversimplified assumption and can be unrealistic in
many practical situations, the naive Bayes classifier exhibits competitive accu-
racy compared to other classifiers, such as logistic regression and decision tree,
especially when the features are not highly correlated. Meanwhile, great efforts
have been made to improve its performance. One common approach is to par-
tially relax the feature independence assumption in naive Bayes, and the result-
ing model is called a semi-­naive Bayes classifier. A review and test of a group
of semi-­naive Bayes classifiers are given in [1], where the authors subdivide the
semi-­naive Bayes classifiers into two groups: the first group uses a new feature
set to construct naive Bayes classifiers, and the second group adds explicitly
directed arcs between nodes to represent their interdependence. Especially, the
new set of features in the first group is achieved by deleting and joining features,
and the interdependence of features in the second group is achieved by introduc-
ing the concept of ‍x ‍-d­ ependence BN classifier, where ‍x ‍is the maximum number
of features allowed to be dependent on each feature in addition to the class.
Obviously, naive Bayes is a 0-­dependence BN classifier. We summarize the typi-
cal semi-­naive Bayes classifiers in these two groups in Table 6.7.
In what follows, we use TAN, which is proposed by Friedman et al. [2], as an
example to show the process of constructing a semi-­naive Bayes classifier. Similar
to the construction of a naive Bayes classifier, all the numerical features and the tar-
get if it is numerical are discretized. The procedure includes the construction of the
qualitative part and quantitative part of a TAN classifier. In particular, the qualita-
tive part refers to the network structure of a TAN classifier and the quantitative part
refers to the CPTs of the nodes. Constructing the qualitative part of the TAN classi-
fier has the following six steps.

Step 1 Specify the prediction target as the class variable and the features
as the attribute variables.
Step 2 The conditional mutual information between each two of the
features (denoted by ‍Ai ‍ and ‍Aj ,‍ and ‍Ai ¤ Aj )‍ given the class variable
(denoted by C ‍ ‍) is calculated (denoted by ‍I(Ai ; Aj |C))‍ to identify the
interdependence of the features.
Step 3 Build a complete undirected graph with attribute variables as the
nodes and the conditional mutual information ‍I(Ai ; Aj |C)‍ as the weight
of the arc between nodes ‍Ai ‍and ‍Aj .‍
Table 6.7   Summary of typical semi-­naive Bayes classifiers

Semi-­naive Bayes classifier Group Basic idea

Backward sequential Group 1 Starting from the full feature set, BSE eliminates the features one by one as long as the
elimination (BSE) classification accuracy of the constructed naive Bayes classifier can be improved according to
leave-­one-­out cross-­validation.
Forward sequential Group 1 Starting from an empty feature set, FSE adds the features that can improve the classification
elimination (FSE) accuracy the most one by one to a naive Bayes classifier according to leave-­one-­out cross-­
validation.
BSE and joining (BSEJ) Group 1 A new Cartesian product of features is formulated in BSEJ by using prediction accuracy as the
merging criterion. Then, BSEJ repeatedly adds or deletes the original and newly constructed
features so as to improve the prediction accuracy as much as possible according to leave-­one-­
out cross-­validation.
Tree augment naive Bayes Group 2 TAN uses an intermediate solution between naive Bayes and unrestricted BN. In TAN, one
(TAN) feature is allowed to depend on at most another feature in addition to the class. Therefore,
TAN is a type of 1-­dependence BN classifier. Conditional mutual information is used to
evaluate feature interdependence.
Naïve-­Bayes Decision Tree Group 2 NBTree is a hybrid approach that combines a naive Bayes classifier and decision tree. The
(NBTree) training data are partitioned using a tree structure and a naive Bayes classifier is constructed in
each leaf node of the tree.
70  Machine learning and data analytics for maritime studies

Step 4 Build the maximum weighted spanning tree by sorting the

weights of the arcs in descending order and then select the arcs from the
largest weight to the smallest weight to link the nodes without forming
a circle. That is, if adding the current largest weighted arc to the nodes
forms a circle, it will not be selected anymore. Then, the availability of
the arc with the second largest weight will be tested. Keep the chosen
arcs and delete the others.
Step 5 Transform the undirected spanning tree into a directed tree by
choosing an arbitrary feature as the root node and setting the direction
of all arcs to other attribute variables to be outward from it.
Step 6 Add the class variable to the tree and arcs from the class variable
to all attribute variables. The construction of the qualitative part of a
TAN classifier is completed.

Especially, in Step 2, the conditional mutual information of attribute variables ‍Ai‍and

A
‍ j‍can be calculated by

P
Ni P P
Nj Nc
P(ai,s0 ,aj,s00 |Cs )
I(Ai ; Aj |C) = P(ai,s0 , aj,s00 , Cs ) log P(a , (6.7)
i,s0 |Cs )P(aj,s00 |Cs )
‍ s0 =1 s00 =1 s=1 ‍
here “‍log ‍” means the logarithmic operation with base 2. ‍Ni ‍, ‍Nj ‍, and ‍Nc ‍ are the
numbers of the possible values of ‍Ai ‍, ‍Aj ‍, and C ‍ ‍, respectively, and ‍ai,s0 ,‍ ‍aj,s00 ,‍ and
‍ s ‍ are specific values of ‍Ai ‍, ‍Aj ‍, and C
C ‍ ‍, respectively. ‍P(ai,s0 , aj,s00 , Cs )‍ is the joint
probability, and ‍P(ai,s0 , aj,s00 |Cs ),‍ ‍P(ai,s0 |Cs )‍ and ‍P(aj,s00 |Cs )‍ are the conditional
probabilities. ‍P(ai,s0 , aj,s00 , Cs )‍should be understood as the proportion of samples
in the whole training set whose states of attribute variables ‍Ai ‍ and ‍Aj ‍ and the
class variable C ‍ ‍ are ‍ai,s0 ,‍ ‍aj,s00 ,‍ and C
‍ s ‍, respectively. Similarly, ‍P(ai,s0 , aj,s00 |Cs )‍
means that among all the samples in the training set whose target is C ‍ s ‍, the pro-
portion of samples whose states of attribute variables ‍Ai ‍and ‍Aj ‍are ‍ai,s0 ‍and ‍aj,s00 ,‍
respectively. The range of ‍I(Ai ; Aj |C)‍is [‍ 0, +1]‍, where a large value of ‍I(Ai ; Aj |C)‍
shows that attribute variables ‍Ai ‍and ‍Aj ‍are more strongly correlated.
After finishing the construction of the qualitative part of the TAN classifier, its
quantitative part is then addressed. The quantitative part has two components: the
marginal distribution and the CPT of each variable, and both can be learned from the
training set. The constructed TAN classifier can then be used to address classification
tasks. An example of developing a TAN classifier to predict whether a new coming
ship has six or more deficiencies is given in Example 6.2.
Example 6.2 The feature processing method is the same as that of Example
6.1. Netica software is used to construct the TAN classifier. The qualitative part
of the TAN classifier is shown in Figure 6.3 and the complete TAN classifier is
shown in Figure 6.4. Figure 6.3 shows that ship type is selected as the root vari-
able, which only has the class variable as its parent. In addition, the last inspec-
tion time is highly related to age and the last deficiency number, while the last
deficiency number is highly related to the last inspection state. It is noted that
the marginal distributions shown in Figure  6.4 are a little different from those
 Bayesian networks  71

Figure 6.3  The qualitative part of the TAN classifier for ship deficiency condition
prediction

shown in Figure 6.2 except for nodes deficiency_no and ship_type, where node
deficiency_no has no parent node and ship_type has only deficiency_no as its
parent node in both figures. This is because the marginal distribution of a node is
calculated based on its CPT, i.e., based on its conditional distributions, in soft-
ware Netica, which would be influenced by its parent nodes. Therefore, when the
parent variable(s) of an attribute variable change(s), it is likely that its marginal

Figure 6.4  The constructed TAN classifier for ship deficiency condition

prediction
Table 6.8   Distribution of last_inspection_time

deficiency_no ship_type 0 to 5 (%) 6 to 10 (%) 11 to 15 (%) 16+ (%) No inspection (%)

0to5 container ship 44.348 38.261 9.565 5.217 2.609
0to5 bulk carrier 35.294 35.294 17.647 5.882 5.882
0to5 tanker 26.087 21.739 21.739 17.391 13.043
0to5 general_cargo/multipurpose 61.539 7.692 15.385 7.692 7.692
0to5 other 9.091 36.364 9.091 27.273 18.182
6+ container ship 44.444 16.667 16.667 11.111 11.111
6+ bulk carrier 47.059 23.529 5.882 17.647 5.882
6+ tanker 22.222 33.333 11.111 11.111 22.222
6+ general_cargo/multipurpose 37.500 31.250 6.250 18.750 6.250
6+ other 45.455 18.182 9.091 18.182 9.091
 Bayesian networks  73

Figure 6.5   The TAN classifier for ship deficiency condition prediction

distribution as shown in the TAN classifier is slightly changed. However, this

does not mean that those nodes’ underlying marginal distributions are changed.
Actually, their marginal distributions are fixed given a dataset. As an illustration,
we present the CPT of node last_inspection_time in Table 6.8.
If the constructed TAN classifier is used to predict the deficiency condition
of the new coming 3-­year-­old tanker ship whose last inspection time is four
months ago with eight deficiencies identified and without detention, the result is
shown in Figure 6.5.
It shows that the coming tanker ship has an 84.4% probability to have 0
to 5 deficiencies and a 15.6% probability to have no less than 6 deficiencies.
Therefore, the tanker ship is predicted to have 0 to 5 deficiencies in the current
PSC inspection.

6.3  BN classifiers

Naive Bayes classifier and semi-­naive Bayes classifier are special cases of BN
classifiers, where no interdependence and partial interdependence of the features
are considered by them, respectively. In general BN classifiers, the interdepend-
ence of each pair of features are presented by directed arcs connecting the attrib-
ute variables. An example of the BN classifier to predict ship deficiency condi-
tions using the same features as Examples 6.1 and 6.2 are shown in Figure 6.6.
The local Markov property [3] is satisfied in BN classifiers, where a node is
conditionally independent of its non-­descendants given its parents. Therefore,
the calculation of joint probability is simplified and is shown in Equation (6.8),
where only each node and the parent node(s) of the node need to be considered.
This property can largely reduce the computation power required to calculate
CPTs and to predict the target, especially in large BNs.
74  Machine learning and data analytics for maritime studies

Figure 6.6  The structure of a BN classifier for ship deficiency condition

prediction

Figure 6.7   Common parent structure in BN classifier

Figure 6.8  ‍‍ V

‍ ‍-­shape structure in BN classifier
 Bayesian networks  75

Figure 6.9   Sequential structure in BN classifier

Figure 6.10  The structure of a BN classifier for ship deficiency condition

prediction

Q
m
P(ck , xi1 , ..., xim ) =P(ck |parent(ck ))  P(xij |parent(xij )), (6.8)
‍ j=1 ‍
Where ‍parent()‍ is the set of the parent node(s) of the node in the parentheses. If
‍parent(xij ) =;,‍ then ‍P(xij |parent(xij )) = P(xij ).‍
There are three typical interdependence relationships of three nodes in the
BN classifier, which are also presented in Figure  6.6: common parent struc-
ture, where at least two child nodes have the same parent node as shown in
nodes ship_type, last_inspection_state, and last_deficiency_no; V ‍ ‍-­shape struc-
ture, where one child node has at least two parent nodes as shown in nodes
last_inspection_state, last_deficiency_no, and deficiency_no; and sequential
structure, where three nodes are connected sequentially as shown in nodes age,
ship_type, and last_deficiency_no. The calculation of the joint probability as
well as the insights generated in each of the three typical structures is shown in
Figures 6.7–6.9.
76  Machine learning and data analytics for maritime studies

Figure 6.11  The constructed BN classifier for ship deficiency condition

prediction

Common parent structure

P(a, b, c) =P(a)p(b|a)p(c|a),
P(a, b, c)
P(b, c|a) = , (6.9)
p(a)
‍ ) P(b, c|a) =P(b|a)P(c|a). ‍
This shows that nodes ‍b‍ and ‍c‍ are conditionally independent given their parent
node ‍a ‍.
V‍ ‍-shape structure
Given the value of node ‍c ‍, nodes ‍a ‍ and ‍b ‍ are interdependent as node ‍c ‍ is
influenced by both of them. However, when the value of node ‍c ‍is unknown, we
have
P
P(a, b) = c P(a, b, c),
P
= c P(c|a, b)P(a)P(b), (6.10)
‍ ) P(a, b) = P(a)P(b), ‍
which shows that nodes ‍a ‍and ‍b‍are independent with each other.
Sequential structure
P(a, b, c)
P(a, c|b) = ,
P(b)
P(b)P(a|b)P(c|b) (6.11)
= ,
P(b)
‍ = P(a|b)P(c|b). ‍
 Bayesian networks  77

which shows that nodes a and c are conditionally independent given the middle
node b.
The BN presented in Figure 6.6 can also be used to predict ship deficiency
condition. The BN classifier learned from the training set is shown in Figure 6.10.
The prediction result of the new coming tanker ship is given in Figure  6.11,
which also suggests that the new coming tanker ship is more likely to have 0 to
5 deficiencies.

References

[1] Zheng F., Webb G.I. ‘A comparative study of semi-­naive bayes methods in
classification learning’. AUSDM05. 2005:141–55.
[2] Friedman N., Geiger D., Goldszmidt M. ‘Bayesian network classifiers’.
Machine Learning. 1997;29(2/3):131–63.
[3] Sebastiani P., Abad M.M., Ramoni M.F. ‘Bayesian networks’. Data Mining
and Knowledge Discovery Handbook. 2009:175–208.
This page intentionally left blank
 Chapter 7
Support vector machine

This chapter first introduces one of the most popular machine learning models for
classification tasks called support vector machine (SVM). Then, kernel trick to
improve its prediction accuracy while reducing the computation burden is discussed.
The extension of SVM to address regression tasks, which is called support vector
regression, is then presented.
For a classification problem on data set ‍D = f(xi , yi ), i = 1, ..., ng‍, where ‍x 2 Rm‍
is the vector of features and ‍y 2 C ‍ is categorical and to be predicted, the basic
idea of the SVM algorithm is to find a proper hyperplane that can accurately sepa-
rate examples in different classes. Consider a binary classification problem, i.e.,
‍y 2 fc1 , c2 g‍, and ‍xi ‍is two-­dimensional, i.e., ‍x = (x1 , x2 )‍. A hyperplane that can dis-
tinguish the two classes of examples, where squares represent examples in posi-
tive class and circles represent examples in negative class, is shown by the lines
in Figure 7.1.
Figure 7.1 shows that hyperplanes that can separate the two classes of exam-
ples are not unique; instead, the number of them can be infinite. Then, how to
choose the most proper one? Intuitively, a hyperplane that has the maximum dis-
tance from the examples in both classes is the best, as it is more tolerant to exam-
ples close to the hyperplane, which are more likely to be misclassified, and thus is
more robust to interference or noises in data. An SVM model aims to find such a
hyperplane, as the hyperplane with the maximum margin can be expected to have
the highest classification accuracy on unseen data, especially for the examples
close to it. Therefore, in Figure 7.1, hyperplane h0 should be the most suitable one.
The dimension of the hyperplane depends on the number of features: if the number
of features is ‍m‍, then the hyperplane is of ‍(m  1)‍ dimensions. In what follows,
we consider the case of binary classification (where 1 indicates a positive example
and ‍ 1‍ indicates a negative example) with two features, where the hyperplane is
a line.

7.1 Hard margin SVM

We first consider the case where examples in the two classes are linearly sepa-
rable, that is, they can be perfectly separated by a straight line. The hyperplane
that can perfectly separate the two classes is a hard margin hyperplane, which is
presented by
80  Machine learning and data analytics for maritime studies

Figure 7.1   Hyperplanes to separate two classes of examples

‍ wT x + b = 0.‍ (7.1)
where ‍w = (w1 , w2 )‍is the normal vector of the hyperplane determining its direction
and ‍b‍ is the offset determining the distance between the hyperplane and the origin.
A hyperplane is determined by the normal vector and the offset, and we denote it by
‍(w, b)‍for simplicity. For example ‍x‍, its distance to ‍(w, b)‍is
|wT x + b|
d(x) = . (7.2)
‍ kwk ‍
For the positive and negative examples in data set ‍ D‍, the following constraints are
satisfied to correctly classify them:
wT xi + b  +1, yi = +1;
(7.3)
‍ wT xi + b  1, yi = 1.‍
Visually, these hyperplanes are shown in Figure 7.2. For all the positive examples,
T
‍w xi + b  +1‍ is satisfied, and the positive examples that are the closest to hyper-
plane h0 satisfy ‍wT xi + b = +1‍. For negative examples, ‍wT xi + b  1‍ is satisfied,
and the negative examples that are the closest to hyperplane h0 satisfy ‍wT xi + b = 1‍.
Examples in both classes that are the closest to the hyperplane are marked in black (the
other examples are in gray) in Figure 7.2. These examples are called support vectors.
The margin is twice the distance between a support vector to the hyperplane, which
can be calculated by
2
= . (7.4)
‍ kwk ‍
According to the basic idea of SVM where the maximum margin should be found,
the problem is then turned to solve the following optimization model:
 Support vector machine  81

Figure 7.2   Hyperplane with the maximum margin and the support vectors

2
max  = . (7.5)
‍ kwk ‍
subject to
wT xi + b  +1, yi = +1, i = 1, ..., n,
(7.6)
‍ wT xi + b  1, yi = 1, i = 1, ..., n.‍
The optimization problem is equivalent to
1
[M0 ] min kwk2 (7.7)
‍ 2 ‍
subject to
‍ yi (wT xi + b)  1, i = 1, ..., n.‍ (7.8)
The optimal hyperplane is denoted by ‍ (w , b ).
‍ Optimization model [‍ M0 ]‍ is a con-
vex quadratic programming problem that can be solved by off-­the-­shelf packages.
Nevertheless, solving its dual problem by introducing the Lagrange multiplier can
largely improve the computational efficiency. By adding the Lagrange multiplier ‍˛i‍,
i‍ = 1, ..., n‍, to each of the constraints in Equation (7.8), the Lagrangian function of
optimization model [‍ M0 ]‍can be written as
82  Machine learning and data analytics for maritime studies

1 X n

[M1 ] L(w, b, ˛) = kwk2 + ˛i (1  yi (wT xi + b)) (7.9)

‍ 2 i=1 ‍
subject to
‍ ˛i  0, i = 1, ..., n‍ (7.10)
where ‍˛ = (˛1 , ˛2 , ..., ˛n ).‍ The optimal ‍ (w , b )
‍can be obtained by setting the par-
tial derivatives of ‍L(w, b, ˛)‍ (denoted by ‍ L‍ for short) to ‍w‍ and ‍b‍ to zero. We first
calculate the partial derivative of ‍L(w, b, ˛)‍to ‍b‍as follows:

@L X n

= ˛i yi = 0,
@b i=1
Xn (7.11)
) ˛i yi = 0.
‍ i=1 ‍
By incorporating Equation (7.11) into objective function (7.9), we can have
1 T X n
X n
Xn

L= w w+ ˛i  ˛i yi wT xi  ˛i y i b
2 i=1 i=1 i=1
Xn
X n (7.12)
1
= wT w + ˛i  ˛i yi wT xi .
‍ 2 i=1 i=1 ‍
Then, the derivative of Equation (7.12) to ‍w‍is
@L X n

=w ˛i xi yi . (7.13)
‍ @w i=1 ‍
and by setting it to zero, we can have
X
n

w= ˛i xi yi . (7.14)
‍ i=1 ‍
By introducing Equations (7.12) and (7.13) to optimization model [‍ M1 ]‍so as to can-
cel out ‍w‍and ‍b‍, the following optimization problem can be obtained:
1 XX
n n
P
n
[M2 ] max˛ ˛i  ˛i ˛j yi yj xTi xj (7.15)
2 i=1 j=1
‍ i=1
‍
subject to
X
n

˛i yi = 0,
i=1
(7.16)
‍ ˛i  0, i = 1, ..., n.‍
The Karush–Kuhn–Tucker condition should be satisfied in [‍ M2 ]‍, as all of the con-
straints are linear. Therefore, the following set of equations holds:
 Support vector machine  83
8̂
ˆ ˛i  0,
<
yi (wT xi + b)  1  0,
ˆ
:̂
‍ ˛i (yi (wT x + b)  1) = 0. ‍
This means that for any example ‍(xi , yi ),‍ we have ‍˛i = 0‍or ‍yi (wT xi + b) = 1.‍ If ‍˛i = 0‍,
it will not appear in Equation (7.12), and thus would not influence the hyperplane.
Otherwise, ‍yi (wT xi + b) = 1‍should be satisfied, which means that the example is on
the maximum margin, i.e., it is a support vector. Therefore, it can be concluded that
the hyperplane obtained by an SVM is only determined by the support vectors while
having nothing to do with the other examples in the training set. After turning the
objective function of optimization model [‍ M2 ]‍to
1 XX X
n n n

min˛ ˛i ˛j yi yj xi xj 
T
˛i . (7.17)
‍ 2 i=1 i=1 i=1 ‍
several efficient algorithms, such as sequential minimal optimization, can be used
to solve [‍ M2 ]‍ so as to obtain the optimal values of ‍ ˛‍ (denoted by ‍ ˛ ‍). Then, the
optimal ‍w‍is
X n

w = ˛i yi xi . (7.18)
‍ i=1 ‍
To calculate the optimal value of ‍b‍, a support vector ‍(xs , ys )‍is first found and com-
bined with the optimal ‍w‍, where we can have
ys (w xs + b ) = 1,
y2s (w xs + b ) =ys ,
(7.19)
y2s = 1,
‍ ) b = ys  w xs .‍
To improve the robustness of the obtained value of b, the set of all support vectors,
denoted by S, can be used to calculate b*:
1 X
b = (ys  w xs ). (7.20)
‍ |S| sS ‍

7.2 Soft margin SVM

In Figure 7.1, examples in the two classes are linearly separable. However, in more
general cases in practice, the two classes of examples are linearly inseparable. That
is, they cannot be separated using a straight line, and thus the hard margin SVM can-
not be applied to the classification task as the constraints cannot be fully satisfied.
An example is shown in Figure 7.3, where the two classes cannot be separated using
a line as they are somewhat “mixed.” One viable approach is to allow an SVM to
make some mistakes in classification while keeping the margin as wide as possible
such that the other points can still be classified correctly. This is the basic idea of a
84  Machine learning and data analytics for maritime studies

Figure 7.3   Examples that are linearly inseparable

soft margin SVM. Then, hyperplane h0 can still be used to separate the two classes,
as presented in Figure 7.4.
Mathematically, the above approach can be realized by allowing some points
(where the set of these points is denoted by ‍ D0 ‍) to violate the constraint,
i.e.,
‍ y0i (wT xi0 + b) < 1, i0 2 D0 .‍ (7.21)
We further denote the violation degree of example i‍0 ‍ of constraint (7.7) by a slack
variable ‍ ‍, ‍ > 0‍, and thus Equation (7.21) can be written as

Figure 7.4   An example of a soft margin SVM

 Support vector machine  85

‍ yi0 (wT xi0 + b)  1  i0 , i0 2 D0 .‍ (7.22)

To unify the format, we also attach the slack variable ‍ ‍ to examples that do not
violate constraint (7.7) while setting ‍ = 0.‍ Meanwhile, optimization model [‍ M0 ]‍is
changed to
1 X n
[M00 ] min kwk2 + C i (7.23)
2
‍ i=1 ‍
subject to
yi (wT xi + b)  1  i , i = 1, ..., n
(7.24)
‍ i  0, i = 1, ..., n. ‍
In objective function (7.23), C‍ ‍is a hyperparameter taking positive value and should
be understood as the penalty on the examples that are wrongly classified. It decides
a trade-­off between maximizing the margin and minimizing the mistakes made. If
‍ ! 1‍, then ‍i ! 0,‍ which means that all the examples are expected to be classi-
C
fied as correctly as possible. If C
‍ ‍takes a finite value, certain errors can be made. If
‍ ‍ is small, less importance is given to classification mistake and more importance
C
is given to maximizing the margin; otherwise, more importance is given to avoid
misclassification at the expense of keeping the margin small. Similar to hard margin
SVM, Lagrange multipliers can be introduced to optimization model [M 0
‍ 0 ],‍ and it is
turned to optimization model [M 0
‍ 1 ]‍
1 X n

[M01 ] L(w, b, ˛, , ) = kwk2 + C i

2 i=1
X
n
X
n (7.25)
+ ˛i (1  i  yi (wT xi + b))  i i
‍ i=1 i=1 ‍
subject to
˛i  0, i = 1, ..., n,
(7.26)
‍ i  0, i = 1, ..., n,‍
where ‍ ˛‍ and ‍‍ are Lagrange multipliers and ‍w‍, ‍b‍, and ‍‍ are model parameters.
Optimization model [M 0
‍ 1 ]‍can be solved similarly to [‍ M1 ]‍.
Soft margin SVM can easily be implemented by the scikit-­learn API [1] in
Python. Here is an example of ship deficiency condition prediction (denoted by
deficiency_state, which is a binary variable with 1 indicating a ship has six or
more deficiencies and 0, otherwise) similar to Example 6.1 in Chapter 6.1, where a
training set with 200 examples and five features, namely, ship age (age), type, last
inspection time (l_ins_time), last deficiency number (l_def_no), and last inspection
state (whether a ship is detained in the last inspection; l_ins_state) are used to
construct a soft margin SVM model.
Example 7.1: As distance needs to be calculated in SVM, all the numerical
features, i.e., age, l_ins_time, l_def_no, and l_ins_state, are processed by min–
max scaler. For the categorical feature type, it is discretized to six features using
86  Machine learning and data analytics for maritime studies

one-­hot encoding: type_bulk_carrier, type_container_ship, type_general_cargo/

multipurpose, type_passenger_ship, type_tanker, and type_other. The target variable
deficiency_state is encoded to a binary variable, where “1” indicates a ship has six or
more deficiencies and “0” otherwise.
We use fivefold cross validation to select the optimal value of hyperparameter
‍ ‍. The search space is set to [‍ 0.001, 0.01, 0.1, 1, 10, 100, 1000].‍ Note that the metric
C
used for parameter selection is ‍F1score‍, as the data set is slightly imbalanced: 46
out of the 200 ships are with six or more deficiencies. The core code is as follows:

‍ ‍

‍ ‍is used to train and test the final soft margin SVM
Then, the optimal value of C
model using the following code:

‍ ‍

The ‍Precision‍, ‍Secall ‍, and F1 score are 0.57, 0.53, and 0.51, respectively.

7.3 Kernel trick

Soft margin SVM is more suitable when the linearly inseparable issue is caused by only
a few examples in classification problems. When the examples in different classes are
interwoven with each other, soft margin SVM might be ineffective. Another approach
to address the problem of linearly inseparable classification problem in SVM is using
kernel trick. The basic idea of kernel trick is to use a transform function to map the
features from the original space to a higher space where they can be linearly separated.
The dot product used in SVM models (i.e., ‍wT x‍) constructed so far on the original
features of each two examples can be viewed as a linear kernel. Moreover, it has been
proven that if the original feature space is of limited dimension, there must be a feature
space of a higher dimension where the examples are linearly separable.
Mathematically, a kernel function can be written as

K(xi , xj ) =h(xi ), (xj )i

(7.27)
‍ = (xi )T (xj ). ‍
 Support vector machine  87

Figure 7.5   A linearly inseparable case

The kernel function in Equation (7.27) can be understood as taking a dot product
(which measures the similarity of the two terms) of the transformed input vectors
by function ‍()‍. The kernel function can easily be incorporated into the original
hard margin SVM model by replacing ‍xi ‍ by ‍(xi )‍ in optimization models [‍ M0 ]‍,
[‍ M1 ]‍, and [‍ M2 ]‍.
An intuitive example to use the kernel trick in SVM to tackle the linearly
inseparable problem is as follows. Suppose we need to separate examples of two
classes presented in Figure 7.5. It is obvious that a straight line cannot separate the
two classes of examples perfectly or nearly perfectly. Instead, a circle can do the job
as shown in Figure 7.6.

Figure 7.6   Using a circle to separate examples in the linearly inseparable case
88  Machine learning and data analytics for maritime studies

Then, a kernel function measuring the similarity of two terms based on whether the
related points are within a circle can be considered. For a point ‍(x10 , x20 )‍in Figure 7.6,
we define a transform function:
p
‍ (x10 , x20 ) = (x210 , x220 , 2x10 x20 ).‍ (7.28)

Then, the kernel function takes the following form:

K((x10 , x20 ), (x100 , x200 )) = h(x10 , x20 ), ((x100 , x200 ))i
(7.29)
‍ = x210 x2100 + x220 x2200 + 2x10 x20 x100 x200 , ‍
which is actually the function form of a circle that can transform the examples into
a 3D space. By doing so, the way to measure the similarity of two terms is changed
from using a dot product (i.e., a linear kernel) to judging whether points are within
a circle by using the new decision boundary generated by a kernel function as is
shown in Figure  7.6. In this way, a linear hyperplane in the 3D space (which is
equivalent to a circle in the 2D space) can perfectly separate the two classes, while
the linear form of the classifier can also be reserved.
Having a closer examination of Equation (7.29), we can find that it is actually
the square of the dot product in the 2D space, i.e.,
K((x10 , x20 ), (x100 , x200 ))
= x210 x2100 + x220 x2200 + 2x10 x20 x100 x200
(7.30)
= (x10 x100 + x20 x200 )2
‍ = h(x10 , x20 ), (x100 , x200 )i2 . ‍
Therefore, the value of the selected kernel function can be calculated directly from the
dot product of two points in the 2D space, and thus the computation burden can be
reduced. The kernel function presented in Equation (7.30) is called a polynomial kernel,
whose degree is 2. The above SVM with kernel trick can be used to address the problem
in Example 7.1 by the scikit-­learn API, which is presented in Example 7.2.
Example 7.2: The features and their processing methods are the same as
those in Example 7.1.1. We also use fivefold cross validation to select the opti-
mal value of parameter C ‍ ‍ with ‍F1score‍ as the metric. The search space is set to
[‍ 0.001, 0.01, 0.1, 1, 10, 100, 1 000]‍. The core code is as follows:

‍ ‍
 Support vector machine  89

It can be found that the optimal value of the penalty term C in this example is
much larger than that in Example 7.1 when linear SVM is used (‍C = 10‍in this case).
This is generally because using polynomial kernel maps the data to a higher dimen-
sion where they are more linearly separable. Then, the tolerance of an example that
is wrongly classified gets smaller, and thus the penalty gets larger. The optimal
value of C‍ ‍is then used to train and test the final soft margin SVM model using the
following code:

‍ ‍

The ‍Precision‍, ‍Recall ‍, and F1 score are 0.57, 0.59, and 0.58, respectively.
Recall that these three metrics for Example 7.1 are 0.57, 0.53, and 0.51, respectively.
Therefore, introducing a polynomial kernel into an SVM increases the prediction
accuracy.
The degree of a polynomial kernel can be any positive integer. The larger
value the degree is, the more complex model will be generated. Therefore, it can
be seen that selecting a proper kernel function is of vital importance to guarantee
the accuracy of an SVM model. This can be a tricky task, as how to accurately map
the original data to a higher dimensional space where they are linearly separable
is unknown. In addition to the linear kernel and the polynomial kernel, some other
common kernel functions are given in Table 7.1.

Table 7.1   Common kernel functions

Kernel function
name Format parameters
Linear kernel ‍K(xi , xj ) =xi xj ‍
T
None
function
Polynomial kernel ‍K(xi , xj ) = (xi xj ) ‍
T d
‍d  1‍, which is the degree of the
function polynomial. A larger ‍d ‍leads to a
more complex SVM model.
Gaussian kernel 
kxi xj k2 ‍ > 0‍, which is the width of the
‍K(xi , xj ) =e 2 2
‍ Gaussian kernel. A smaller ‍ ‍leads
to an SVM model that is more
sensitive to the distance between
each pair of points, and thus is more
likely to overfit the data.
Laplacian Kernel kxi xj k The same as Gaussian kernel.
‍K(xi , xj ) =e
 
‍
Sigmoid kernel ‍K(xi , xj ) = tanh(ˇxi xj + )‍ ‍tanh ‍is the hyperbolic tangent function,
T

‍ˇ > 0,‍ ‍ < 0‍.

90  Machine learning and data analytics for maritime studies

7.4 Support vector regression

SVM is designed for classification problems. It can be extended to address regres-

sion problems by conducting proper revisions, and the model is called support vec-
tor regression or SVR for short. In standard ML models for regression tasks, if the
predicted output and the real output are not equal, their difference will be counted in
the loss function such as MSE and MAE. Nevertheless, SVR aims to find a line with
a certain width to fit the data, such that the error of the examples within the line with
certain width is not considered when calculating the value of the loss function, while
only the examples outside the line with certain width will be considered. An illustra-
tion of SVR where only one feature is considered is shown in Figure 7.7. The line
of the SVR model developed to predict the output of the examples is ‍f(x) =wx + b,‍
and the width is ‍ 2‍. Thus, the error of the examples falling within the area bounded
by lines ‍f(x) + ‍ and ‍f(x)  ‍ will not be accounted into the calculation of the loss
function, while the others will be accounted for.
Mathematically, an SVR model can be formulated as follows:
1 Xn

[M3 ] min kwk2 + C l (f(xi )  yi ). (7.31)

w,b 2
‍ i=1 ‍
where C ‍ ‍is the penalty term, and ‍l (f(xi )  yi )‍is the loss function given the value of
‍‍which takes the following form:
8
< 0, if |f(xi )  yi |  
l (f(xi )  yi ) =
: |f(xi )  yi |  , otherwise.
‍ ‍
Then, two slack variables, ‍i1‍ and ‍i2,‍ are introduced to linearize ‍l (f(xi )  yi ),‍ and
model [‍ M3 ]‍is turned to

Figure 7.7   An illustration of an SVR model

 Support vector machine  91

1 X n

[M03 ] min kwk2 + C (i1 + i2 ) (7.32)

w,b,i1 ,i2 ,i=1,...,n 2
‍ i=1 ‍
subject to
f(xi )  yi   + i1 , i = 1, ..., n,
yi  f(xi )   + i2 , i = 1, ..., n,
(7.33)
i1  0, i = 1, ..., n,
‍ i2  0, i = 1, ..., n. ‍
Similar to optimization model [M 0
‍ 1 ],‍
Lagrange multipliers can be introduced to model
[M 0 ] so as to obtain its dual problem. Then, the format of an SVR model can be writ-
‍ 3‍
ten as
X n

f(x) = (˛i1  ˛i2 )xi x + b. (7.34)

‍ i=1 ‍
where ‍˛i1‍and ‍˛i2‍are the Lagrange multipliers for example i‍ ‍, i‍ = 1, ..., n‍, for the first
two constraints in Equation (7.33). Examples with ‍˛i1  ˛i2 ¤ 0‍ are the support
vectors of the SVR, which are the examples lying outside of the line with width ‍ 2‍.
Obviously, they are a subset of the whole training set. Similar to SVM, kernel trick
can also be applied in SVR to improve model performance.
SVR is also implemented in the scikit-­learn API. The following example is to
predict the number of deficiencies of a ship (i.e., a continuous output) using the same
features as well as their decoding methods in Example 7.2.
Example 7.3: We first use fivefold cross validation to select the most suitable
kernel function from “linear,” “poly” (i.e., polynomial kernel), “rbf” (i.e., Gaussian
kernel), “sigmoid,” the degree value if “poly” kernel is used, and the penalty term C ‍ ‍.
The core code is as follows:

‍ ‍

Then, the optimal kernel and its degree and the value of the penalty term C
‍ ‍are
used to train and test the final SVR model using the following code:
92  Machine learning and data analytics for maritime studies

‍ ‍

The ‍MSE ‍, ‍ MAE ‍, and ‍ MAPE ‍are 11.1886, 2.3153, and 0.7085, respectively.

Reference

[1] Pedregosa F., Varoquaux G., Gramfort A., et al. ‘Scikit-­learn: machine learning
in python’. Journal of Machine Learning Research. 2011;12:2825–30.
 Chapter 8
Artificial neural network

This chapter aims to introduce a widely-­used machine learning model for both
regression and classification tasks called artificial neural network (ANN). Basic
concepts of ANNs are first covered, and then typical algorithms and tricks for ANN
training are introduced. Finally, deep learning models, as an important type of neural
network, are briefly discussed.

8.1 The structure and basic concepts of an ANN

Suppose we have a data set ‍D = f(xi , yi ), i = 1, ..., ng‍, where ‍xi 2 Rm ‍is the vector
of features and ‍yi 2 R ‍ is the output. The structure of a typical ANN model to
predict yi (denoted by ‍yO i ‍, i‍ = 1, ..., n ‍) is presented in Figure 8.1. It contains three
layers: an input layer, a hidden layer, and an output layer. Especially, the input
layer contains ‍m‍ nodes to receive the ‍m‍ feature values of the examples, which
can either be categorical or numerical, in addition to a bias node. The output
layer gives the final predicted target ‍yO i ‍. For a specific problem, the structure of
the input and output layers is fixed. In contrast, the hidden layer can be much
more flexible: an ANN can have one or more hidden layers, and the number of
nodes in one hidden layer is also flexible. In particular, the number of hidden
layers and the number of nodes contained by each of them depend on the spe-
cific problem and the data structure. In Figure 8.1 one hidden layer with ‍ K + 1‍
nodes is contained. Weight is attached to each arrow connecting two nodes in
consecutive layers in Figure  8.1. The direction of the arrows shows the data
flow direction in an ANN model, which is opposite to the direction of the error
flow which will be covered later.
One node in the neural network is also called a neuron. The details of a neu-
ron in Figure  8.1 are shown in Figure  8.2. A neuron is an elementary unit of an
ANN model, which can be viewed as a mathematical function receiving one or more
inputs from an example or from the neurons in the preceding layer. For the neuron
shown in Figure 8.2, it receives inputs from a1, a2, and a3 with weights w1, w2, and w3
attached. The weighted sum of the inputs, which is denoted by ‍t ‍, is first calculated
by ‍t = w1 a1 + w2 a2 + w3 a3 + b‍, where ‍b‍is the bias. Then, ‍t ‍is passed to an activation
function ‍f ‍to generate the output of this neuron, which is denoted by ‍ z ‍, i.e., ‍z = f(t).‍
In particular, activation functions deciding whether the neurons should be activated
94  Machine learning and data analytics for maritime studies

Figure 8.1   An illustration of a typical ANN model

play a fundamental role in ANNs as they introduce nonlinearity (i.e., nonlinear

transformation) into the network. Nonlinearity guarantees the universal approxima-
tion property of ANNs: “Standard multilayer feedforward networks with as few as
one hidden layer using arbitrary squashing functions are capable of approximating
any Borel measurable function from one finite dimensional space to another to any
desired degree of accuracy, provided sufficiently many hidden units are available”
[1]. Popular activation functions are presented in Figures 8.3–8.6.

Figure 8.2   An illustration of a neuron in ANNs

 Artificial neural network  95

12 Applications of machine learning and data analytics models

TableFigure
10.10 8.3  
Association rules mined
An illustration with min_Conf
of Sigmoid f(x) = 1+e=1x1.5
0.7 and ‍min_Lift
activation=function: ‍

Association rulethe structure andConfidence

Based on Lift
mathematical representation of a neuron, the layers
in anD11
D10, ANN shown in Figure  0.81
) D07 8.1 can be mathematically1.58 presented as follows. As
statedD10
D04, above,
) D07the input layer of an ANN model receives
0.80 1.57the input from the feature
D07,
values. ) D10
D11Then, they serve as the0.70 1.61 and the weighted sum of
input to the hidden layers,
the inputs to neuron ‍bk , k = 1, ..., K ‍denoted by tk, is calculated by
P
m
tk = wm0 k xm0 , k = 1, : : : , K. (8.1)
‍ m0 =1 ‍
Then, the output of node  ‍k, k= 1, ..., m1 ˇ
P Kˇ + 1‍is ˇˇ
distman (xi , xj ) =xi  xj 1 = xim0  xjm0 .
zk = f(tk ), k = 1, ..., K, m0 =1
P (8.2)
=zK+1 =xm1.
m1
‍ 0 A0 xm0 B0 . ‍
m0 =1
Note that, for simplicity, we set the bias terms ‍xm+1 = 1‍ in the input layer and
‍zK+1 = 1‍in the hidden layer. As the bias term is connected by its weights, which can
be adjusted, to the neuron(s) in the consecutive layer, the value of the bias term itself
10.3 FP-growth algorithm
can be set to an arbitrary nonzero constant and is usually set to be 1. The outputs of
nodes in the hidden layer also serve as the input to the output layer, which is
The main difference between the FP-growth algorithm and the Apriori algorithm is
P
k+1
in the tcontext
y = vof generating all large item sets, that is, finding all item sets with
k zK . (8.3)
Sup(I
‍ min_Sup‍ , while the procedure of generating all association rules is the
n0 )  k=1
same in both algorithms. FP-growth algorithm is proposed by Han et al. in 2000 [4]
and is believed to be an improvement of Apriori as it is much faster. In Apriori, data
are stored in set structure, and candidate large item sets should be generated, the num-
ber of which can be large. Moreover, the whole database needs to be traversed multi-
ple times to check the Support values of the candidate item sets in a brute-force man-
ner, and traversing the whole data set heavily reduces the computational efficiency. In
the FP-growth algorithm, data are stored in a tree structure called a frequent pattern
96  Machine learning and data analytics for maritime studies

x x
Figure 8.4   An illustration of Tanh activation function: ‍f(x) = eexe
+ex ‍

In regression tasks where the output is numerical, the weighted sum of the outputs
given by the hidden layer is used directly as the final prediction in the output layer,
that is

‍ yO = ty .‍ (8.4)
In classification tasks where the output is categorical, the continuous weighted sum
input to the output layer is further discretized. In binary classification problems,
sigmoid function is often used, where the predicted value given by the output layer
is the probability of being a positive class:
1
yO = . (8.5)
‍ 1 + ety ‍
In multiclass classification problems, the output is multi-­dimensional, and a Softmax
function is often used to process the outputs, which turns absolute values into prob-
abilities. We do not discuss the details of this case here. Then, the predicted output
of the output layer can be uniformly expressed as

‍ yO = g(ty ).‍ (8.6)

0
Therefore, the overall relationship between the input features ‍xm0 ,‍ ‍m = 1, ..., m‍, and
the output ‍yO ‍can be presented by
 Artificial neural network  97

Figure 8.5  An illustration of ReLU (short for rectified linear units) activation
function: ‍f(x) = 0‍if ‍x < 0‍and ‍f(x) =x ,‍ otherwise

yO = g(ty )
P
k+1
= g( vk zk )
k=1
P
k
(8.7)
= g( vk f(tk ) + vK+1 )
k=1
P k P
m+1
= g( vk f( wm0 k xm0 ) + vK+1 ).
‍ k=1 m0 =1 ‍

8.1.1 Training of an ANN model

In the above analysis, the output of one layer serves as the input to its directly con-
secutive layer. Then, the output of the last layer, i.e., the output layer, gives the final
prediction ‍yO .‍ The cost function is an evaluation of the loss brought about by the
predicted output ‍yO ‍ with regard to the real output ‍y ,‍ which is used to evaluate the
prediction error. In regression problems, the cost function is usually half of the MSE
between the real and the predicted output values. In binary classification problems,
cross entropy is a popular cost function. That is, if yi is numerical, the cost function
of this example can be calculated by
98  Machine learning and data analytics for maritime studies

Figure 8.6  An illustration of leaky ReLU activation function: ‍f(x) = 0.2x ‍if
‍x < 0‍and ‍f(x) =x ‍, otherwise

1
L(yi , yO i ) = (yi  yO i )2. (8.8)
‍ 2 ‍
If yi is binary, the cost function of this example can be calculated by
‍ L(yi , yO i ) =[yi (log(Oyi )) + (1  yi )(log(1  yO i ))].‍ (8.9)
The training process of an ANN model in one round is to adjust the weights connect-
ing the neurons, with the goal to minimize the cost function: the smaller the value of
the cost function, the closer the predicted and actual outputs. The most popular ANN
learning algorithm is called backpropagation, which is a way of computing gradients
of expressions through recursive application of chain rule, such that the current error
in the output layer can be reversely passed to each layer, and the weights can be
adjusted accordingly to minimize the error. We use the case of regression problem
as an example to show the working mechanism of the backpropagation algorithm.
We first calculate the gradients of the weights connecting the hidden layer and
the output layer. Assume sigmoid function is used as the activation function, i.e.,
1 0
‍f(x) = 1+ex .‍ Then, its derivative can be calculated by ‍f (x) =f(x)(1  f(x)).‍ Denote
the current training example by ‍(xi , yi ).‍ For neuron ‍k, k = 1, ..., K ‍in the hidden layer
whose weight is denoted by vk, its connection with the output layer can be repre-
sented by
 Artificial neural network  99
8̂
1
ˆ
ˆ L = (yi  yO i )2
ˆ
ˆ 2
ˆ
<
yO i = ty (8.10)
ˆ
ˆ
ˆ
ˆ XK+1
ˆt = vk zk .
:̂ y
‍ k=1 ‍
Then, the gradient of ‍ L‍with respect to vk, denoted by ‍ı vk ‍for short, can be calculated
by the following chain rule:
@L
ıvk =
@vk
@L @Oyi @ty (8.11)
=
@Oyi @ty @vk
‍ = (Oyi  yi )zk . ‍
where all the terms involved are known. The gradient of weight ‍wm0 k ‍ connecting
neuron ‍m0 ‍ in the input layer and neuron ‍k ‍ in the hidden layer can be calculated in
a similar way. The connection of neuron ‍m0 ‍ and the prediction error can be repre-
sented by
8̂
1
ˆ
ˆ L = (yi  yO i )2
ˆ
ˆ 2
ˆ
ˆ
ˆ yO = t
ˆ
ˆ
ˆ i y
ˆ
ˆ
ˆ
ˆ X
K+1
<
ty = vk zk
(8.12)
ˆ
ˆ k=1
ˆ
ˆ
ˆ
ˆ zk = f(tk )
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ X
m+1
ˆ
ˆ tk =
:̂ wm0 k xm0 .
‍ m0 =1 ‍
Then, the gradient of ‍ L‍with respect to ‍wm0 k ‍, denoted by ‍ıwm0 k ‍ for short, can be
calculated by the following chain rule
@L
ıwm0 k =
@wm0 k
@L @Oyi @ty @zk @tk (8.13)
=
@Oyi @ty @zk @tk @wm0 k
‍ = (Oyi  yi )vk f(tk )(1  f(tk ))xm0 .‍
where all the terms involved are known. ‍ı vk ‍ and ‍ıwm0 k ‍ are then used to update
the weights connecting the hidden layer and the output layer and those connect-
ing the input layer and the hidden layer, respectively. To reduce the problem
of over-­fitting, learning rate, which is denoted by ‍ ,‍ is proposed to control the
speed of weight updating. Then, after this round of learning, the weight con-
necting neuron ‍k ‍, ‍k = 1, ..., K ‍and the output layer is updated to
100  Machine learning and data analytics for maritime studies

‍ vQ k = vk  ıvk ,‍ (8.14)

and the weight connecting neuron ‍m0 ‍, ‍m0 = 1, ..., m‍, and neuron ‍k ‍, ‍k = 1, ..., K ‍, is
updated to
‍ wQ m0 k = wm0 k  ıwm0 k .‍ (8.15)
By doing so, the current prediction error is fed back throughout the entire net-
work, and the weights can be adjusted with the aim to minimize the prediction
error. Note that in one round of the above updating using standard backpropa-
gation, only one example, i.e., ‍(xi , yi )‍, is used to calculate the gradient of the
weights. As only a single random example is used, standard backpropagation
might lead to the problem of local minima, and the updating results might be
unstable. An alternative is to use the whole training set in one round to calculate
the gradients, which is also called accumulated error backpropagation. However,
in many tasks, when the accumulated error is small, it can be very slow to further
reduce it. Under this condition, standard backpropagation is more suitable. To
reach a trade-­off between standard backpropagation and accumulated error back-
propagation, a few examples are selected randomly, instead of a single example
or the entire training set, for each iteration for weight updating. The set of exam-
ples is called a mini batch, and the number of examples in one mini batch is called
batch size. The process of weight updating is iterated several rounds, where the
number of rounds is denoted by “‍epoch‍.”

8.1.2 Hyperparameters in an ANN model

Hyperparameters in an ANN model can be divided into two categories: one is
related to network structure, and the other is related to learning algorithm. The
number of hidden layers and the number of neurons contained in each of the hid-
den layers are two main types of hyperparameters to control the complexity of
the network structure in ANNs. Generally speaking, more hidden layers or more
neurons contained in one hidden layer increase model complexity, and thus might
lead to the problem of over-­fitting. One trick to reduce the problem of over-­fitting
is called dropout. It is a regularization approach that can be applied to input and
hidden layers. Dropout means for an input or a hidden layer, temporarily remov-
ing a certain proportion (denoted by ‍p )‍ of neurons from the layer, including the
neuron itself as well as the connections with its preceding (if any) and following
layers. Then, the outputs of this hidden layer are scaled by multiplying each of
the outputs by (1‍  p ).
‍ This approach is widely used in deep neural networks
(DNNs). Besides, the types of activation functions of the neurons are also hyper-
parameters that need to be decided before ANN training, and they can have a
large impact on model performance.
Hyperparameters regarding the learning algorithm mainly include the learning
rate (i.e., ‍ ),
‍ the number of rounds (i.e., ‍epoch),‍ and batch size. In particular, ‍ ‍
controls the speed of weight updating. If ‍ ‍is too small, the speed of learning would
be slow and a larger value of ‍epoch‍ might be needed. If ‍ ‍ is too large, the optimal
 Artificial neural network  101

values of the weights might be surpassed in the updating process. If the value of
‍epoch‍ is too large, the ANN model developed might learn the data too well, lead-
ing to the problem of over-­fitting. In contrast, if ‍epoch‍is too small, the problem of
under-­fitting might occur. To find a proper value for ‍epoch,‍ a validation set that is
independent of the training set should be used to test the performance of the tem-
porary ANN model constructed: if the cost function on the validation set decreases
moderately or even increases, the training should be stopped as the problem of over-­
fitting is highly likely to occur. This trick is also called “early stopping.” Finally,
batch size is highly dependent on the size of the whole data set and the network
structure. Common batch size is 1, 2, 4, 16, 32, 64, 128, and 256. In addition, to
overcome the problem of over-­fitting, regularization on the weights in the network
is introduced to the cost function shown in Equation (8.8):

1 X
n
X
m
X
k+1
X
k+1

LQ =  L(Oyi , yi ) + (1  )( w2m 0k + v2k ), (8.16)

n
‍ i=1 m0 =1 k=1 k=1 ‍
where ‍ ‍ is a hyperparameter that needs to be tuned, which balances the trade-­off
between model accuracy and network complexity. Note that regularization is not
imposed on the bias term, as it is usually a parameter with a fixed and small value
(e.g., 1). All the hyperparameters mentioned above should be tuned by a hold-­out
validation set or based on cross validation and grid search.
ANN models for regression tasks are implemented in scikit-­learn API [2],
which is named “MLPRegressor” under the class of “neural_network.” We use the
following example to show the procedure of developing an ANN model to predict
ship deficiency number.
Example 8.1. We first randomly divide the whole data set containing 1 991
examples to a training set with 1 500 examples and a test set with 491 examples.
The numerical features considered are GT, age, last_inspection_time, and last_defi-
ciency_no. The categorical features considered are flag_performance, RO_perfor-
mance, company_performance, last_inspection_state, and ship type. Especially, for
flag_performance, white is encoded to 0, gray is encoded to 1, black is encoded
to 2, and unknown is encoded to 3. For RO_performance, high is encoded to 0,
medium is encoded to 1, and unknown is encoded to 2. For company_performance,
high is encoded to 0, medium is encoded to 1, low is encoded to 2, very low is
encoded to 3, and unknown is encoded to 4. For last_inspection_state, if a ship has
no last inspection record in the database of Tokyo MoU, it is decoded to ‍ 1‍ (this
also applies to numerical features last_inspection_time and last_deficiency_no). If a
ship is detained in the last inspection, it is encoded to 1. If a ship is not detained in
the last inspection, it is encoded to 0. Ship type is encoded using one-­hot encoding
the same as the examples in Chapter 7.
We fix the number of hidden layers as one and the number of neurons contained
as 10. All the nodes in the hidden layer use ReLU function as the activation function.
The following hyperparameters are tuned by fivefold cross validation on the train-
ing set: learning rate (learning_rate_init in scikit-­learn API), batch size (batch_size
Table 8.1   Popular deep learning models

Model name Main structure Applications

Convolutional neural network (CNN) Three main types of layers are contained: Supervised and unsupervised: image processing,
convolution, pooling, and fully-­connected object detection.
layers.
Recurrent neural network (RNN) RNNs have cyclic connections in the structure Supervised: speech recognition, handwriting
that employ recurrent computations to recognition, and text recognition where
sequentially process the input data. The edges sequences and patterns are involved.
in an RNN are fed into the next time step
instead of into the next layer in the same time
step.
Long short term memory network (LSTM) LSTM is a variant of RNN that can handle Supervised: time-­series prediction, speech or
long-­term dependencies. Therefore, an LSTM video recognition, music composition, and
network has the same structure as an RNN, pharmaceutical development.
except that the neurons in the hidden layer
of the RNN are replaced by memory blocks.
The basic component of an LSTM network is
called a memory block, which contains a cell
and three gates.
Deep belief network (DBN) DBNs integrate restricted Boltzmann machines Supervised and unsupervised: image
(RBMs) with deep feedforward neural recognition, video recognition, motion-­
networks (D-­FFNNs), where RBMs form the capture data processing.
input unit and D-­FFNNs form the output unit.
Generative adversarial network (GAN) GANs have two components, namely, a Supervised and unsupervised: Image generation,
generator and a discriminator, and they can human face generation, cartoon character
create new data that resemble the training generation, image transformation, and
data by learning from them. astronomical image improvement.
 Artificial neural network  103

in scikit-­learn API), and epoch (max_iter in scikit-­learn API). The core code is as
follows:

‍ ‍
Then, the optimal learning rate init, batch size, and max iter are used to train
andtest the final ANN model using the following code:

‍ ‍
The ‍MSE ‍, ‍ MAE ‍, and ‍ MAPE ‍are 14.3317, 2.5591, and 0.6437, respectively.

8.2 Brief introduction of deep learning models

With the improvement of computation resources, training more complex ANN mod-
els where adequate storage and computational power are needed becomes possible.
With the accessibility of big data, the problem of over-­fitting in these complex mod-
els can be largely reduced. The most common way to improve the complexity of an
ANN model is to increase the number of its hidden layers: the deeper a network, the
more complex a model. The reasons are intuitive: as the number of hidden layers
becomes larger, more weights between consecutive layers are added, and a larger
number of activation functions associated with the neurons can be involved. As a
result, their joint efforts can significantly increase the complexity and capacity of an
ANN model. An ANN model with a large number of hidden layers is called a DNN,
or a deep learning model. The term “deep learning” describes a family of learning
algorithms rather than a single algorithm to develop DNN models. Popular deep
learning models are summarized in Table 8.1.
104  Machine learning and data analytics for maritime studies

References

[1] Hornik K., Stinchcombe M., White H. ‘Multilayer feedforward networks are
universal approximators’. Neural Networks. 1989;2(5):359–66.
[2] Pedregosa F., Varoquaux G., Gramfort A., et al. ‘Scikit-­learn: machine learning
in python’. Journal of Machine Learning Research. 2011;12:2825–30.
 Chapter 9
Tree-­based models

This chapter aims to introduce several machine learning models based on a tree
structure called the decision tree, which are widely believed to be among the most
popular methods for both classification and regression tasks. The basic structure and
concepts as well as the tree-­growing algorithms of a single decision tree are first
introduced. As a single decision tree is prone to over-­fitting, ensemble models con-
sisting of a certain number of decision trees are developed. Random forest-­based on
bagging and gradient boosting decision trees based on boosting will be introduced
as the representatives of ensemble models on decision trees.

9.1 Basic concepts of a decision tree

As the name suggests, a decision tree, which we call DT for short, uses a tree-­like model
for decision making. A DT consists of nodes and directed edges, where a set of examples
is contained in a node, and a directed edge means splitting a node into consecutive nodes.
An example of a DT is shown in Figure 9.1. The orange node at the top is a root node,
which contains the whole data set used to construct the current DT. The green nodes are
leaf nodes that are not split any more and give the final prediction. The blue node is the
internal node, which is to be further split. The directed edges in the DT show the process
of node splitting: a parent node at the tail of a directed edge is split into two child nodes at
the head following some criteria, which is usually a selected feature or a (feature, value)
pair. The term “Splitting” here means that the examples contained by the parent node are
separated into two complementary and disjoint child nodes, and the splitting criterion is
the feature or the (feature, value) pair. The examples with feature values less than or equal
to the threshold value are split to the left child node, and others to the right child node. The
splitting aim is to make the two child nodes become “purer.” The term “purer” here means
that the examples contained by one node are more similar to each other: in classification
problems, the examples in one node are expected to be in the same class; in regression
problems, the targets of examples in the same node should be as close as possible. Finally,
the output of a leaf node is also determined by the problem. In general, in classification
problems, the class of the majority of the examples contained within it is used as the out-
put; in regression problems, the average target of all the examples contained is the output.
The depth of a tree is one more than the number of splits from the root node to the deepest
leaf node. The depth of the tree in Figure 9.1 is 3.
106  Machine learning and data analytics for maritime studies

Figure 9.1   The structure of an example DT

Figure 9.1 shows that a DT is constructed in a recursive manner: from the root

node, the nodes are split following the same rule until one node cannot be split any
more (e.g., only one example is contained by it or the targets of all the examples
contained by it are identical) or some preset criteria are reached (e.g., the tree has
reached the maximum depth, or the number of examples contained by a leaf node of
the current node, if it is to be split, will be less than the minimum number of exam-
ples required to be in a leaf node). Then, when using a constructed decision tree to
predict the output of an unseen example, the prediction process follows a set of “if-­
then” rules from the root node to a certain leaf node. In the process of decision tree
construction, the most important task is how to select the optimal splitting criteria
to split a node. The strategies used to find the optimal splitting criteria are different
in classification and regression tasks and in different tree construction algorithms,
which will be introduced in the following sections in detail.

9.2  Node splitting in classification trees

Three algorithms are popular to construct classification DTs: iterative dichotomizer

3 (ID3), C4.5, and classification and regression tree (CART).

9.2.1 Iterative dichotomizer 3 (ID3)

In ID3, a node in a decision tree is repeatedly (iteratively) divided (dichotomize) into two
or more child nodes from top to bottom in a greedy manner. The term “greedy” means
that for each node, the best splitting criteria at the present moment for the current node is
selected, while it may not be the optimal of the whole tree. ID3 is based on entropy and
information gain in information theory. Especially, entropy is used to measure the degree
of randomness in the information being processed. The higher the entropy, the harder to
draw any conclusion from the information. Denote by ‍ X ‍a discrete random variable with
‍I ‍values which has the following probability distribution
 Tree-­based models  107

Figure 9.2  The relationship between entropy and probability in binary

classification problems

‍ P(X = xi ) =pi , i = 1, ..., I.‍ (9.1)

Then, the entropy of random variable ‍ X ‍ is
P I
H(X) = pi log2 pi . (9.2)
‍ i=1 ‍
If ‍pi = 0‍, ‍pi log pi ‍is set to 0. The entropy of ‍ X ‍only depends on its distribution, and
thus Equation (9.2) can be written as Equation (9.3):
P I
H(p) = pi log2 pi . (9.3)
‍ i=1 ‍
In particular, in binary classification problems where ‍ I = 2‍, the probability of one
class is ‍p‍and that of the other class is 1‍  p‍. The relationship between entropy ‍H(p)‍
and probability ‍p‍can be shown as follows:
Figure  9.2 shows that when ‍p = 0‍ or ‍p = 1‍ we have ‍H(p) = 0‍, which means
that there is no uncertainty in the random variable. When ‍p = 0.5‍we have ‍H(p) = 1‍,
which means that the uncertainty in the random variable is at the maximum. For two
random variables ‍ X ‍ and ‍ Y ‍, conditional entropy ‍H(X, Y)‍is the uncertainty of ‍Y ‍ when
‍X ‍is known, and can be presented as follows
P I
H(Y|X) = pi H(Y|X = xi ), (9.4)
‍ i=1 ‍
where ‍pi = P(X = xi ), i = 1, ..., I .‍ Based on the concepts of entropy and conditional
entropy, information gain (IG) is proposed as a statistical property that measures how
much uncertainty can be reduced by introducing another random variable as the condi-
tion, that is,
108  Machine learning and data analytics for maritime studies

‍ G(Y, X) =H(Y)  H(Y|X).‍ (9.5)

The concept of IG can be used as the criterion to find the optimal split of a node in a
DT for classification, where Equation (9.5) can be understood as how much uncer-
tainty regarding the categorical target in the training set can be reduced if a feature
is introduced as the condition. The larger the IG value of a feature, the more suitable
the feature is for splitting the node, as the subsequent nodes are purer. This is the
basic idea of the ID3 algorithm for DT construction.
ID3 is applicable to classification problems where all the features are categori-
cal. In the process of DT construction using ID3, from splitting the root node, the
IG of each candidate feature for splitting the current node (where each feature
value results in a child node) is calculated, and the feature leading to the largest
IG is used as the feature to split the node. Especially, the child nodes of the cur-
rent node are obtained by using all the values of the selected feature to split the
current node, and the number of resulting child nodes is equal to the number of
values the selected feature has. Note that for a numerical feature, the values of
the feature may be infinite, while there might be only one example with a specific
feature value. Consequently, it is believed that ID3 is not able to deal with numeri-
cal features.
The above process can be mathematically presented as follows. Given training set
‍D = f(xi , yi ), i = 1, ..., ng‍ for a classification problem, where the target has a total of ‍ K ‍
possible values and one value is denoted by ‍ck , k = 1, ..., K ‍. For example, in the ship deten-
tion prediction problem, ‍ K = 2‍ and ‍c1 = 0‍and ‍c2 = 1‍. 1 indicates that a ship is detained
and 0, otherwise. C ‍ k ‍is the number of examples in class ck. For a feature ‍x‍(e.g., ship type)
with ‍J ‍candidate values, i.e., the set of values of ‍x‍is fx
‍ 1 , ..., xj , ..., xJ g‍, at the root node, ‍ D‍
can be divided into ‍J ‍subsets with each value of ‍x‍as a child node (i.e., examples with the
same value of ‍x‍are divided into one child node), which are denoted by ‍D1 , ..., Dj , ..., DJ ‍.
Denote the set of examples belonging to class ck in ‍Dj‍by ‍Djk ‍. For a node other than the
root node, denote the set of examples contained in it by ‍D0 , D0 2 D‍, the set of examples in
subset ‍j‍by ‍Dj0,‍ the number of examples in ‍ D0‍belonging to class ck by C 0
0 0
‍ k ,‍ and the set of
examples belonging to class ck in ‍Dj ‍by ‍Djk .‍The entropy of ‍ D ‍ is
0

P
K
C0k C0k
H(D0 ) = .
|D0 |
log2 |D0 | (9.6)
‍ k=1 ‍
Then, the conditional entropy of feature ‍x ‍over data set ‍ D0 ‍ is
P
J |D0 | P
J |D0 | P
K
0|
|Djk 0|
|Djk
H(D0 |x) = |Dj0 | H(Dj0 ) = |Dj0 | k=1 |Dj0 |
log2 |Dj0 |
. (9.7)
‍ j=1 j=1 ‍
Then, the IG can be calculated by
‍ G(D0 , x) =H(D0 )  H(D0 |x).‍ (9.8)
Feature ‍x ‍ leading to the maximum value of among all candidate features is used


as the optimal feature to split the current node, where the examples of the same
value for ‍x‍are split to one child node. This means that a total of ‍J ‍child nodes are
 Tree-­based models  109

generated. Then, the child nodes generated are split following the above steps. A
node will not be split any more if its entropy is below a preset threshold.
ID3 is an intuitive algorithm based on information theory, which is easy to
understand. However, it has some obvious drawbacks, which are listed as follows:

•• ID3 cannot properly process continuous feature, as a continuous feature usually

has too many possible values and each of the subsequent child nodes may only
contain one example. Therefore, continuous features cannot be directly used by
the ID3 algorithm.
•• IG gives an absolute value which is highly related to the number of feature
values: IG gets large accordingly if the number of feature values increases.
For example, ship type has many possible values such as container ship, bulk
carrier, heavy lift cargo ship, tanker ship, chemical or LNG ship, multipur-
pose ship, passenger ship, fishing vessel, tug, liquefied natural gas tanker, and
nuclear fuel vessel, among others. In contrast, ship RO performance only has
three possible values, namely high, medium, not listed in our data set. If IG is
used as the criteria to select the best feature for splitting, it would be biased
toward feature ship type as this feature has more values, and thus it is more
likely to be selected.
•• Missing values of the features cannot be processed by ID3.
•• There is no pruning process in a DT constructed by ID3 when it was first pro-
posed, and thus might lead to the problem of over-­fitting.

9.2.2 C4.5
The above drawbacks of ID3 can be addressed by C4.5 algorithm in the following
ways, which is developed by Quinlan [1]:

•• For a continuous feature ‍x ‍ with ‍J ‍ values (where ‍J ‍ can be close to ‍n‍ espe-
cially for non-­integer values), dichotomy is used to discretize the features. First,
rank the values of feature ‍x ‍ in ascending order denoted by fx 1 j J
‍ , ..., x , ..., x g.‍
Then, there will be a total of ‍J  1‍ split points that can divide the values of
feature ‍x ‍ into two categories. One candidate split point is denoted by ‍xt ‍, and
t xj +xj+1 t 0
‍x = 2 , j = 1, ..., J 0 1.‍ Based on ‍x ‍, the current data set ‍ D ‍can be dividedt
into two categories: ‍Dxt‍with examples with feature ‍x ‍less than or equal to ‍x ,‍
and ‍D0xt+‍ with feature ‍x ‍ larger than ‍xt ‍. Then, the continuous feature ‍x ‍ can be
treated as a binary feature.
•• Mainly aimed for a categorical feature ‍x ‍with ‍J ‍values, information gain ratio,
or IGR, is used as the criteria to find the best split feature in C4.5 algorithm. The
IGR of feature ‍x ‍to data set ‍ D0 ‍can be calculated by
G(D0 , x)
GR (D0 , x) = , (9.9)
‍ Hx (D0 ) ‍
110  Machine learning and data analytics for maritime studies

PJ |Dj0 | |Dj0 |
where Hx (D0 ) = j=1 |D0 |
log2 |D0 | ‍ is the intrinsic value of feature ‍x ‍. The
‍
larger number of values of ‍J ‍, the larger value of ‍Hx (D0 )‍. Hence, it can be used to
reduce the influence of the number of feature values on the IG value calculated.
•• Two issues regarding missing values need to be addressed: one is how to find
the best split feature when some features have missing values, and the other is
after selecting the best split feature, how to deal with the examples with miss-
ing feature values. The basic idea to address the above two issues is to attach
different weights to the examples. Readers are referred to Reference 1 for more
detailed information.
•• Two tree pruning strategies can be adopted: prepruning and postpruning. In
prepruning, some limitations called tree growing termination conditions are set
to restrict tree growing, which aim to prevent the tree from learning the training
data too well by limiting its size. Common tree growing termination conditions
include:

1. max_tree_depth, which is the maximum depth of the tree.

2. min_samples_leaf, which is the minimum number of examples that should be
contained by a leaf node of the tree.
3. min_gain, which is the minimum IG or IGR required for splitting a node.

In postpruning, a complete tree is first constructed, and then two leaf nodes from
the same parent node are “drawn back” to their parent node, i.e., assume that the parent
node is not split. Then, the performance of the DT with leaf nodes drawn back is tested
on a hold-­out validation set. If the prediction performance can be improved, the two
leaf nodes involved are to be pruned, and they are reduced to their parent node.

9.2.3 Classification and regression tree (CART)

CART is a binary tree where each node is split into two child nodes following a
given splitting criterion: examples satisfying the criterion are split to the left child
node, and the other examples are split to the right child node. As the name suggests,
CART can be used for both classification and regression tasks. In this section, we
introduce how to apply CART algorithm to construct DT for classification tasks. In
a classification tree constructed by CART, Gini index is used as the criterion to find
the best split. For data set ‍D = f(xi , yi ), i = 1, ..., ng‍where ‍yi 2 fc1 , ..., ck , ..., cK g‍and
ck is a possible target value. The number of examples in class ck is denoted by C ‍ k ‍.
The probability of an example to be of class ck is pk. Then, the Gini index of the
probability distribution is
P
K P
K
Gini(p) = pk (1  pk ) = 1  p2k . (9.10)
‍ k=1 k=1 ‍
For binary classification problem, denote the probability of an example to be of one
class is ‍p‍, then the Gini index of the probability distribution can be calculated by

‍ Gini(p) = 2p(1  p).‍ (9.11)

 Tree-­based models  111

The Gini index of data set ‍ D‍can be calculated by

PK
‍ Gini(D) = 1  k=1 ( C|D|k )2 .‍ (9.12)
In CART classification tree where the nodes are divided in a binary manner, feature
‍x ‍as well as its value ‍xj ‍is used to divide the current node containing data set ‍ D0 ‍ into
two mutually exclusive nodes containing data set ‍D01‍and ‍D02,‍ respectively. If feature
‍x ‍is continuous, ‍D01‍contains examples with feature value less than or equal to ‍xj ‍and
0 j 0
‍D2‍contains examples with feature value larger than ‍x ‍. If feature ‍x ‍is categorical, ‍D1‍
j 0
contains examples with feature value ‍x = x ‍and ‍D2‍contains examples with feature
value ‍x ¤ xj .‍ Then, the Gini index of data set ‍ D0 ‍if feature value pair ‍(x, xj )‍is used
for node splitting can be calculated by
|D0 | |D0 |
Gini(D0 , x, xj ) = 10 Gini(D01 ) + 20 Gini(D02 ). (9.13)
‍ |D | |D | ‍
Similar to the entropy of data set ‍ D0‍, the Gini index of ‍ D0‍shows its degree of uncer-
tainty: the larger value of Gini(D 0 ), the higher degree of uncertainty in ‍ D0 ‍. Therefore,
‍ ‍
to split a node containing data set ‍ D0‍ in a classification tree based on CART, the
feature value pair leading to the minimum Gini(D 0 , x, xj ) is selected for node splitting.
‍ ‍
CART is able to process missing values in features, and it is based on finding
surrogate splits. We do not cover the details here. Readers are referred to [2] for more
illustration. A DT constructed using CART can be pruned to reduce over-­fitting. The
prune process starts from pruning the leaf nodes to their parent node until coming to
the root node. Then, a set of sub-­trees can be formulated in the pruning process, which
is denoted by fT‍ 0 , T1 , ..., TN g‍. Then, a hold-­out validation set or cross validation should
be used to find the best sub-­tree leading to the minimum loss function:
‍ C˛ (Tn ) =C(Tn ) + ˛|Tn |,‍ (9.14)
‍ n )‍ is the prediction error of sub-­tree T
where C(T ‍ n‍ on the validation set or in cross
validation, |T
‍ n |‍is the number of leaf nodes in the sub-­tree used to represent tree com-
plexity, and ‍ ˛‍is a hyperparameter taking non-­negative value to achieve a trade-­off
between tree prediction error and tree complexity.
The overall procedure of constructing a classification tree is shown in
Algorithm 1.

‍ 3
112  Machine learning and data analytics for maritime studies

An improved version of CART is implemented by scikit-­learn API [3] to deal

with classification problems. We use the same features as those in Example 9.1.
Chapter 9 to predict whether a ship has more than 5 deficiencies in the follow-
ing Example 9.1.
Example 9.1. Recall that the whole data set containing 1991 examples is
randomly divided into a training set with 1500 examples and a test set with
491 examples. The numerical features are GT, age, last_inspection_time, and
last_deficiency_no. The categorical features are flag_performance, RO_per-
formance, company_performance, last_inspection_state, and ship type. The
feature encoding method is the same as that in Example LABEL:eg:ANN_reg.
Note that as the features as well as their values are used to find the best split
directly without calculating distance or similarity, there is no need for feature
scaling.
We use Gini index as the node splitting criteria, and we tune the values of two
hyperparameters to control tree complexity: max_depth and min_samples_leaf. As the
training set is unbalanced, where only 381 ships have more than 5 deficiencies among
all the 1500 ships contained, ‍F1 score‍is used as the metric for hyperparameter tuning.
The core code is as follows:

‍ ‍
 Tree-­based models  113

Then, the optimal max_depth and min_samples_leaf are used to train and test
the final classification DT model using the following code:

‍ ‍

The Precision, Recall, and F1 score are 0.66, 0.64, and 0.65, respectively.

9.2.4 Node splitting in regression trees

When applying DTs to address regression tasks, CART is the most popular algo-
rithm for tree construction, and we only cover CART in this section. Given a training
data set ‍D = f(xi , yi ), i = 1, ..., ng‍where yi is continuous, MSE is used as the criterion
to split one node in a regression tree in CART. Starting from the root node, the tree is
constructed in a greedy manner: to split one node containing data set ‍ D0 ‍, all features as
well as their values are enumerated to form the feature value pair denoted by (feature,
value) or ‍(x, xj )‍. Then, examples with the feature less than or equal to the threshold
value are split to the left child node, and the other examples are split to the right child
node. That is, the example set of the left child node is ‍D01 (x, xj ) =fi = 1, ..., n|xij  xj g,‍
j
and the example set of the right child node is ‍D02 (x, xj ) =fi = 1, ..., n|xij g.‍ The output
of one child node is the average targets ofP the examples contained in that node. That
is, the output of ‍D01 (x, xj )‍is c1 = 0 1 j 0 j
i2D10 (x,xj ) yi , and the output of ‍D2 (x, x )‍is
‍ |D1 (x,x )| ‍
P
c = 1 0 j y . The best split pair is the one that leads to the minimum sum
‍ 2 |D02 (x,xj )| i2D2 (x,x ) i ‍
of MSE of the two child nodes, that is
" #
P P

x )
(x , = j
arg min 2
(yi  c1 ) + 2
(yi  c2 ) . (9.15)
‍ ‍
0 i2D02 (x,xj )
x2fx1 ,...,xJ g,xj 2fx1 ,...,xJ g i2D1 (x,xj )

After a regression tree based on CART is constructed, tree pruning can be conducted
similar to the classification tree based on CART to reduce over-­fitting. The overall
procedure of constructing a regression tree using CART is shown in Algorithm 2.

‍
114  Machine learning and data analytics for maritime studies

An improved version of CART is implemented by scikit-­learn API [3] to deal

with regression problems. We use the same features and prediction target as
those in Example 9.1 of this chapter to predict the number of deficiencies of a
ship in the following Example 9.2.
Example 9.2. We use MSE as the node splitting criteria, and we tune the
values of two hyperparameters to control tree complexity: max_depth and
min_samples_leaf, similar to that in classification DT. The core code is as
follows:

‍ ‍

Then, the optimal max_depth and min_samples_leaf are used to train and test
the final regression DT model using the following code:

‍ ‍
 Tree-­based models  115

The ‍MSE ‍, ‍ MAE ‍, and ‍ MAPE ‍are 18.4986, 2.8218, and 0.6946, respectively.
DT is a popular machine learning method to address both classification and
regression problems. The advantages and disadvantages of DTs are summarized as
follows.

Advantages:

•• DTs are easy to understand and intuitive. They can also be visualized. Decision
rules can be summarized from a constructed DT.
•• Not many data as well as feature preprocessing procedures are needed. For
example, feature normalization and scaling are not needed, and some DT con-
struction algorithms can deal with missing values.
•• DTs can easily be extended to deal with multi-­output problems.

Disadvantages:

•• DTs can easily be over-­fitting, especially when the tree constructed is very
complex.
•• DTs can be unstable, which means that they can easily be influenced by small
variations in the data, especially some extreme feature values.
•• The prediction given by DTs for regression problems is not smooth or
continuous.
•• Learning an optimal DT is known to be NP-­hard. Therefore, all the above popu-
lar DT construction methods are in a heuristic manner, where optimal decisions
in a local manner (i.e., for each node in the tree) are made*.

We summarize the features and properties of the common algorithms for DT

construction as follows in Table 9.1.

9.3 Ensemble learning on tree-based models

The above disadvantages of DTs can be mitigated by ensemble learning, a

technique in machine learning model to combine a certain number of indi-
vidual learners to reach a final decision, where the basic structure is shown in
Figure 9.3.
The individual learners contained by an ensemble learning model can be
identical in terms of model type (i.e., homogeneous) or distinct (i.e., heteroge-
neous). In the former case, the individual learners are called base learners, and

* 
Constructing global optimal DTs is also studied by researchers, where readers are refered to [4–6]
Table 9.1   Common DT construction algorithms

Name of the Output data Input data Splitting Tree structure Processing of Pruning
algorithm criterion missing values
ID3 Categorical Categorical Information gain Polytree (i.e., a tree No None
data data node can have
more than two
child nodes)
C4.5 Categorical Categorical Information gain Polytree Yes, based on Prepruning and postpruning based on
data and ratio assigning nodes drawn back (i.e., pessimistic
numerical weights to error pruning)
data examples
CART Categorical Categorical Classification: Binary tree (i.e., a Yes, based on Prepruning and postpruning which is
and and Gini index; tree node only surrogate first generating all possible sub-­trees
numerical numerical Regression: has two child splits and then finding the best one (i.e.,
data data MSE (mean nodes) cost-­complexity pruning)
squared error)
or MAE (mean
absolute error)
 Tree-­based models  117

in the latter case, they are called component learners. In this section, we mainly
focus on how to develop an ensemble model consisting of a certain number of
DTs constructed by CART algorithm as base learners to address classification
or regression tasks. Actually, based on Hoeffding inequality, as the number of
base learners contained by the ensemble model, i.e., ‍T ‍, increases, the error rate
decreases dramatically and will be close to zero eventually [7]. Meanwhile, it
should also be noted that the above analysis is based on an important assump-
tion: the error of the base learners should be independent with each other.
However, “good” while “independent” base learners are hard to find, because
all the base learners are developed to address the same problem, and thus they
cannot be independent, especially under the condition that the data set size is
limited while certain accuracy should be achieved. Therefore, the core steps of
developing a strong ensemble model include how to develop “divergent and
good” base learners that can capture different aspects of the data accurately,
and how to produce a strong ensemble model based on these base learners. To
achieve this, two strategies are widely used, namely bagging and boosting.

9.3.1 Bagging
As mentioned in section 9.2.4, one disadvantage of DTs is that they are very sensi-
tive to the training set: a slightly different training set can result in a totally different
DT, especially for training sets with extreme points. Therefore, to produce diver-
gent base learners in an ensemble model, different training sets generated from the
original training set can be used. At the same time, as the dimension of the original
training set is fixed and limited, to ensure that the performance of a base learner is
not too bad, the number of examples contained by a training set for a base learner
should be the same as the original training set. One viable approach is to draw
samples of the same size as the original training set from the whole training set and
construct machine learning models on these samples. Bootstrapping aggregating
(bagging) is such a widely used approach to develop ensemble learning method
in a parallel manner. Bagging is based on bootstrap sampling: for a training set ‍ D‍
containing ‍n‍ examples, to formulate a sample training set ‍D O t ,‍ a random example
is first selected and place to ‍DO t ,‍ and then it is placed back to ‍ D ‍. The above process
is repeated ‍n‍ times, and then the sample training set ‍D O t ‍ with ‍n‍ examples can be
formulated. As the examples are selected with replacement, it is highly likely that
there are duplicate examples in ‍D O t ‍ while some examples in ‍ D ‍are not contained
by ‍DO t .‍ Actually, only about 63.2% of the examples in ‍ D ‍are chosen in ‍D O t ‍ when
the number of examples contained in training set ‍ D ‍ is sufficiently large [8]. After
repeating the above process ‍T ‍ times, ‍ T ‍sample training sets, ‍D O 1‍to ‍DO T ,‍ can be for-
mulated, and can be used to train ‍T ‍DTs. To ensemble the outputs of the ‍T ‍DTs in
the final ensemble model, voting strategy is adopted in classification problems, i.e.,
the target of the majority DTs is used as the final output. In case of tie, a random
output in tie is chosen. In regression problems, the average of the targets of all the
DTs is used as the final output.
118  Machine learning and data analytics for maritime studies

Figure 9.3   An illustration of ensemble learning model

A famous tree ensemble method, called random forest or RF for short, is an

extension of bagging method based on DTs where one more layer of randomness
is introduced to the training process of a DT. The only difference between bag-
ging based on DTs and RF is the candidate features used to split a node in a DT:
all the features are considered to find the optimal feature value pair in bagging,
while only a subset of all the features are considered to find the optimal feature
value pair in RF. Denote the number of features used to split a node of a DT in the
RF by ‍d ‍ and recall that data set ‍ D‍ has ‍m‍ features, the recommended value for ‍d ‍
is ‍log2 m‍ [9]. Therefore, the randomness in RF comes from two aspects: samples
from bootstrap sampling used to construct each DT and a subset of features used
to split each node. The characteristics and working mechanism of an RF model is
shown in Figure 9.4.
RF is implemented by scikit-­learn API [3] to deal with both regression
and classification tasks. We use the following Example 9.3 and Example 9.4 to
show how to call scikit-­learn API to predict ship detention and ship deficiency

Figure 9.4   An illustration of an RF model

 Tree-­based models  119

number in PSC. The features and target of Example 9.3 for classification are the
same as Example 9.1, and those of Example 9.4 for regression are the same as
Example 9.2.
Example 9.3. In addition to the hyperparameters in DT, RF has more hyperpa-
rameters to control the whole structure of the model. The number of DTs contained,
i.e., ‍T ‍, is called n_estimators in scikit-­learn, the number of features to consider when
looking for the best split for a node is called max_features. Similar to the DT con-
structed to predict ship detention, we also use Gini index as the node splitting criteria
of each DT. We tune a total of four hyperparameters: two from the DT to control tree
complexity: max_depth and min_samples_leaf, and the other two from the ensemble
model n_estimators and max_features. As the training set is unbalanced, where only
381 ships have more than 5 deficiencies among all the 1500 ships contained, ‍F1 score‍
is used as the metric for hyperparameter tuning. The core code is as follows:

‍ ‍

Then, the optimal n_estimators, max_features, max_depth, and min_samples_leaf

are used to train and test the final classification DT model using the following code:

‍ ‍

The ‍Precision‍, ‍Recall ‍, and F1 score are 0.68, 0.63, and 0.65, respectively.
Example 9.4. We use MSE as the node splitting criteria, and we tune the values of
four hyperparameters to control the structure of RF and tree complexity: n_estimators,
max_features, max_depth, and min_samples_leaf, similar to that in classification RF.
The core code is as follows:
120  Machine learning and data analytics for maritime studies

‍ ‍

Then, the optimal n_estimators, max_features, max_depth, and min_samples_leaf

are used to train and test the final regression RF model using the following code:

‍ ‍

The ‍MSE ‍, ‍ MAE ‍, and ‍ MAPE ‍are 14.2134, 2.5567, and 0.6212, respectively.

9.3.2 Boosting
The base learners are independently constructed and are ideally independent of each
other in bagging; therefore, they can be constructed in parallel. In contrast, the base
learners, which are also called weak learners, are strongly dependent on each other
in boosting. The weak learners are constructed one by one, which aims to constantly
improve (boost) the performance of the current model developed by paying more
attention to the examples with larger prediction errors by attaching higher weights
to them or changing the prediction target. Then, after developing ‍T ‍weak learners for
the boosting model, they are combined linearly to produce the final output. In boost-
ing, the number of weak learners ‍T ‍is also called the maximum number of iterations.
In this section, we introduce two boosting methods: adaptive boosting or Adaboost
and boosting tree, where the former is based on changing the examples’ weights and
the latter is to learn the error of the base models constructed so far.

9.3.2.1 Adaboost
The weak learner of Adaboost is the decision stump, which is a DT with only
one split on the root node. In each round of weak learner construction, the
 Tree-­based models  121

weights of the examples are changed according to their prediction error in

Adaboost, and the examples with weights attached are used to train the weak
learner in the next round. To be more specific, consider a binary classifica-
tion problem with a training set denoted by ‍D = f(xi , yi ), i = 1, ..., ng,‍ where
‍yi 2 f1, 1g.‍ The procedure of constructing an Adaboost model is shown in
Algorithm 3.

‍ ‍
122  Machine learning and data analytics for maritime studies

There are some further explanations of Algorithm 3.

•• In step 1, all the examples are assigned identical weights to train the first (an
initial) weak learner. Then, the weights of the examples will be updated accord-
ing to the performance of the current weak learner.
•• Examples that are wrongly predicted will be assigned higher weights as shown
by how ‍w t+1,i ‍ is calculated. Therefore, these examples need more attention in
the next round of learning. Therefore, it can be seen that in the learning process
of Adaboost, the examples are not changed while their weights, i.e., the distri-
bution of the examples, are changed.
•• From Equation (1.16), it can be seen that the error of round ‍t ‍ is the sum of
weights of the examples that are wrongly classified, which indicates the
relationship between distribution ‍D t ‍ of the data set and the weak learner
‍ t (x) .‍
G
•• The weight of weak learner G ‍ t (x) ,‍ i.e., ‍˛ t ‍ is important of this weak
learner in the final classification model. It is shown in Equation (1.17)
that when ‍e t  12 ,‍ ‍˛ t  0 ,‍ and as et decreases ‍˛ t ‍ increases. This shows that
the more accurate a weak learner is, the more important role it plays in
the final prediction model. In contrast, when ‍e t > 12 ,‍ we have ‍˛ t < 0 .‍ It is
not reasonable to have a weak learner whose weight is less than zero. To
address this issue, one approach is to change the predicted outputs to their
inverse numbers, and the error rate should become smaller than ‍12 .‍ Another
approach is to discard the current weak learner as its performance is
too bad.
•• The sign of ‍f Adaboost
combined
(x)‍ determines the predicted class of example ‍x ‍, and the
absolute value of ‍f Adaboost
combined
(x)‍shows the reliability of the prediction.
•• Regularization can be applied in Adaboost in Equation (1.18) of Step 7 by introduc-
ing learning rate denoted by ‍v ‍, ‍v 2 (0, 1),‍ which is used to control the model updat-
ing speed. Then, Equation (1.19) is changed to ‍f Adaboost t
(x) =f Adaboost
t1
(x) + v˛ t G t (x).‍
The larger value ‍v ‍is, the faster the model learns.

The following Example 9.5 is an example to use Adaboost for ship risk prediction
using the same features and target as those in Example 9.3.
Example 9.5. As decision stump is used as the weak learner, there is no hyper-
parameter used to control the complexity of a DT contained by an Adaboost. Only
two hyperparameters are tuned to control the learning performance of the whole
Adaboost model: n_estimators, i.e., ‍ T ‍, which is the number of weak learners con-
tained, and learning_rate, i.e., ‍v ‍ to control model learning speed. As the training
set is unbalanced, where only 381 ships have more than five deficiencies among all
the 1500 ships contained, ‍F1 score‍is used as the metric for hyperparameter tuning.
The core code is as follows:
 Tree-­based models  123

‍ ‍

Then, the optimal n_estimators and learning_rate are used to train and test the
final classification DT model using the following code:

‍ ‍

The ‍Precision‍, ‍Recall ‍, and F1 score are 0.69, 0.66, and 0.67, respectively.
Adaboost is widely used for classification tasks, and it can also be used for
regression tasks. When used for regression, the main idea and procedure are identi-
cal to the situation when it is used for classification, except for the equations used
to calculate the performance of a weak learner in Equation (9.16) of Algorithm 3
(which is determined by the regression algorithm adopted), the weight of the exam-
ples in ‍ D‍, and the weight of each weak learner in Equation (9.17). In addition,
theoretically, any ML model can be used as weak learner under the framework of
Adaboost, and the most popular ones are DTs and ANNs. It should also be noted that
Adaboost can be very sensitive to outliers or extreme examples, as these examples
will be attached with large weights in model constructing as they are more likely to
be badly predicted. Consequently, the performance of Adaboost is badly influenced.

9.3.2.2 Boosting tree
Boosting tree is also called gradient boosting decision tree, or GBDT for short,
which can be used for classification and regression tasks. It is based on CARTs and
uses forward-­stagewise additive modeling where the model is fit by minimizing
a loss function averaged over the training data, i.e., to use the prediction error in
the previous round of learning as the new target in the next round of weak learner
124  Machine learning and data analytics for maritime studies

learning. The above idea can be summarized as “learn from the prediction error so
far,” and it is achieved by changing the prediction target in the current round to the
negative gradient value of the loss function in the last round. The procedure of using
GBDT for a regression task is shown in Algorithm 4.

‍ ‍

When using MSE as the loss function in this above task, the negative gradient calcu-
lated in Equation (1.21) is ‍yi  G t1 (xi ), i = 1, ..., n,‍ which is the difference between the
real and current predictions for example i‍ ‍in the training set. This difference is also called
residual and is denoted by ri. Therefore, it can be seen that the learning objective of each
round in GBDT for regression is to minimize the residual of the current prediction, and
the final output is calculated by summing the outputs of all rounds of learning. The fol-
lowing Example 9.6 is used to show how to use the scikit-­learn API [3] to develop a
GBDT model for regression to predict the number of deficiencies a ship has.
Example 9.6. In addition to the hyperparameters in DT (i.e., max_depth and
min_samples_leaf), GBDT for regression has more hyperparameters to control the
 Tree-­based models  125

construction of the whole ensemble model. Similar to Adaboost, it has n_estimators

and learning_rate to control the learning process of the whole model. We use MSE
as the node splitting criteria, and we tune the values of four hyperparameters to con-
trol the structure and complexity of GBDT: n_estimators, learning_rate, max_depth
and min_samples_leaf. The core code is as follows:

‍ ‍

Then, the optimal n_estimators, learning_rate, max_depth and min_samples_

leaf are used to train and test the final regression RF model using the following
code:

‍ ‍

The ‍MSE ‍, ‍ MAE ‍, and ‍ MAPE ‍are 16.4603, 2.6489, and 0.6671, respectively.
The basic idea of using GBDT for classification is similar to applying GBDT for
regression. Consider a GBDT model for binary classification. The main difference
is that the output of GBDT for classification is discrete, and thus the prediction error
cannot be directly used as the target in the next round of learning like GBDT used
for regression. This problem can be solved by using exponential loss, which is simi-
lar to the case of Adaboost, or using log loss, which is similar to the case of logistic
regression. In this section, we discuss the case of using log loss for classification
tasks by GBDT. Log loss aims to minimize the difference between the predicted
probability and the real probability of the examples. The log loss function for clas-
sification can be presented by

‍ L(y, yO ) = ln(1 + exp(yOy)),‍ (9.25)

Table 9.2   Comparison of boosting and bagging based on DTs

Name of Representative models Type of base Relationship of base Learning Combining weak Learning rate
algorithm learners learners objective of learners
each round of
learning
Bagging RF CART Independent, and can Real target Averaging in No
be paralleled regression
and voting in
classification
Boosting Adaboost and GBDT Adaboost: Dependent, and Adaboost: real Adaboost: weights; Yes
decision cannot be target; GBDT: GBDT: additive
stump; GBDT: paralleled negative manner
CART gradient
 Tree-­based models  127

where ‍y 2 f1, 1g‍. The procedure of using GBDT for classification task is shown
in Algorithm 5.

‍ ‍

Regularization can be used in GBDT to reduce the problem of over-­fitting.

For each base learner based on CART, pruning of each tree is a viable approach.
Another way is to use sub-­sampling without replacement to construct one weak
128  Machine learning and data analytics for maritime studies

learner, with the aim to reduce the over-­fitting of each weak learner. For the whole
GBDT model, one possible way is to introduce learning rate ‍v ‍, ‍v 2 (0, 1)‍like that in
Adaboost to control the model learning speed, and then Equation (9.23) is changed
PJ t
to G ‍ t (x) =G t1 (x) +Pv j=1 c tj I(x 2 R tj )‍ in Algorithm 4 and (9.29) is changed to
Jt
‍ t (x) =G t1 (x) + v j=1 c tj I(x 2 R tj )‍in Algorithm 5.
G
We compare boosting and bagging models based on DTs in Table 9.2.

References
[1] Quinlan J.R. C4.5: programs for machine learning. San Mateo, CA: Morgen
Kaufmann Publishers; 2014.
[2] Hastie T., Tibshirani R., Friedman J.H. ‘The elements of statistical learning’ in
The Elements of Statistical Learning: Data Mining, Inference, and Prediction.
Vol. 2. New York, NY: Springer; 2009.
[3] Pedregosa F., Varoquaux G., Gramfort A, et al. ‘Scikit-­learn: machine learning
in python’. Journal of Machine Learning Research. 2011;12:28–2825.
[4] Bertsimas D., Dunn J. ‘Optimal classification trees’. Machine Learning.
2017;106(7):82–1039.
[5] Aghaei S., Gómez A., Vayanos P. ‘Strong optimal classification trees’. [arXiv
preprint arXiv:210315965] 2021.
[6] Verwer S., Zhang Y. ‘Learning optimal classification trees using a binary lin-
ear program formulation’. Proceedings of the AAAI Conference on Artificial
Intelligence. 2019;33:32–1625.
[7] Hoeffding W. ‘Springer’. [In: The Collected Works of Wassily Hoeffding]
Probability Inequalities for Sums of Bounded Random Variables. 1994::26–409.
[8] Breiman L. ‘Bagging predictors’. Machine Learning. 1996;24(2):40–123.
[9] Breiman L. ‘Random forests’. Machine Learning. 2001;45(1):5–32.
 Chapter 10
Association rule learning

Association rule learning is a rule-­ based and unsupervised machine learning

method, which is widely used to discover interesting relations between items (i.e.,
records) in a database and also to explore how and why these items are connected.
A widely known application of association rule learning is to analyze the purchased
items in market basket transactions, which aims to identify what goods are often
bought together in one transaction so as to adjust sales strategies to increase sales.
For example, a famous story is that on Friday nights, the sales of diapers and beer
were correlated in Walmart: a bottle of beer was often bought when diapers were
bought. The explanation was that working men were asked to pick up diapers on
their way home from work, and they would also buy a bottle of beer for themselves
at the same time. Based on this finding, Walmart put the shelves of these two goods
close to each other on Friday nights, and the sales volume of both increased greatly.
In the above example, diapers and beer are correlated in the following way: diapers
‍ ‍beer, meaning that if diapers are bought, beer is highly like to be bought in the
!
same transaction. This is a basic form of association rule we are going to cover in
this section. Actually, association rule learning is widely used to mine the relations
of items from transaction databases, and the rules generated are used to guide the
activities of personalized product recommendation, shelf placement, combined cou-
pon dispatching, and bundle sales in the retailing industry.

10.1 Large item sets

To introduce how association rules are generated from a given database, we first
introduce some basic concepts. Denote a set of ‍N ‍ items by ‍I = fi1 , ..., iN g‍. An
item set is a subset of ‍ I ‍, and an item set containing ‍n0 ‍ items (where ‍n0 2 [1, ..., N]‍)
is also called an ‍n0 ‍item set and is denoted by ‍In0 ‍. A database consisting of ‍ M ‍ records,
where each record is an item set, is denoted by ‍D = ft1 , ..., tM g‍. Define the event of
observing the occurrence of a particular item set ‍In0 ‍ by ‍E(In0 )‍, which means that
all the items in ‍In0 ‍ are observed in one record. We further define ‍P(E(In0 ))‍ as the
proportion of the ‍ M ‍records that have all the items in ‍In0 ‍, which can also be inter-
preted as the probability of the occurrence of event ‍E(In0 )‍. It should also be noted
that a record that includes all the items in item set ‍In0 ‍ can also include items not in
‍In0 ‍but in ‍ I ‍. The probability of ‍E(In0 )‍is also called the ‍Support ‍of item set ‍In0 ‍, that
130  Machine learning and data analytics for maritime studies

is, ‍P(E(In0 )) = Sup(In0 )‍, and ‍P(E(In0 )) 2 [0, 1]‍. A larger value of ‍P(E(In0 ))‍ indicates
that the more frequently item set ‍In0 ‍occurs in ‍ D‍. In order to define a large item set
denoted by ‍In0 ‍ which frequently occurs in the database, we define the minimum
threshold of ‍Support ‍for an item set to be a large item set by ‍min_Sup‍. That is, if and
only if ‍In0  I ‍and ‍In0 ¤ ;‍is a large item set, we have ‍Sup(In0 )  min_Sup.‍
A rule is generated by dividing a large ‍n0 ‍ item set, i.e., ‍In0 ‍ and ‍n0  2‍, into
two mutually exclusive and non-­empty item sets ‍Ij ‍and ‍Ik ,‍ with ‍Ij [ Ik = In0 .‍ A rule
can be generated from ‍Ij ‍ to ‍Ik ‍ in form ‍Ij ! Ik .‍ To determine whether rule ‍Ij ! Ik ‍
is an association rule denoted by ‍Ij ) Ik ,‍ two indicators are further introduced:
Confidence
‍ ‍and ‍Lift ‍of ‍Ij ! Ik ‍is calculated by
P(E(Ij ) \ E(Ik ))
Conf(Ij ! Ik ) = = P(E(Ik )|E(Ij )), (10.1)
‍ P(E(Ij )) ‍
and Conf(I
‍ j ! Ik ) 2 [0, 1].‍ The larger value Confidence
‍ ‍ is, the more likely that the
items in ‍Ik ‍appear given that the items in item set ‍Ij ‍appear. ‍Lift ‍of ‍Ij ! Ik ‍is calcu-
lated by
P(E(Ij ) \ E(Ik )) P(E(Ik )|E(Ij ))
Lift(Ij ! Ik ) = = , (10.2)
‍ P(E(Ij ))  P(E(Ik )) P(E(Ik )) ‍
and ‍Lift(Ij ! Ik ) 2 [0, +1)‍ presents the influence of the occurrence of event ‍E(Ij )‍
on event ‍E(Ik )‍, which is the ratio of the probability of the occurrence of event ‍E(Ik )‍
under the condition that event ‍E(Ij )‍occurs and the probability that event ‍E(Ik )‍occurs
unconditionally in the database. This can be interpreted as how the occurrence of
event ‍E(Ij )‍can increase/decrease (i.e., lift) the occurrence of ‍E(Ik )‍. To be more spe-
cific, if ‍lift(Ij ! Ik ) 2 [0, 1),‍ the occurrence of ‍E(Ij )‍decreases the probability of the
occurrence of ‍E(Ik )‍. If ‍Lift(Ij ! Ik ) 2 (1, +1),‍ the occurrence of ‍E(Ij )‍increases the
probability of the occurrence of ‍E(Ik )‍. If ‍Lift(Ij ! Ik ) = 1,‍ the occurrence of ‍E(Ij )‍
has no influence on the occurrence of ‍E(Ik )‍, that is, ‍E(Ij )‍and ‍E(Ik )‍are independent.
It is also interesting to find that as event ‍E(Ik )‍ acts as the denominator to calculate
‍Lift ‍ of rule ‍Ij ! Ik ,‍ if ‍P(E(Ik ))‍ is large, meaning that the occurrence probability of
event ‍E(Ik )‍ is high, the value of ‍Lift(Ij ! Ik )‍ would be reduced. This shows that a
frequently occurring event would have less contribution to generating association
rules compared to rare events.
The concept of association rule is defined based on Confidence ‍ ‍ and ‍Lift ‍ of a
rule as follows:
Definition 10.1. Given a large item set ‍In0‍with ‍n0  2‍and two non-­empty, mutu-
ally exclusive, and complementary item sets ‍Ij ‍and ‍Ik ‍, i.e., ‍Ij ¤ ;,‍ ‍Ik ¤ ;,‍ ‍Ij \ Ik = ;,‍
and ‍Ij [ Ik = In0,‍ and given the minimum thresholds of Confidence
‍ ‍ and ‍Lift ‍ denoted
by ‍min_Conf ‍and ‍min_Lift ,‍ respectively, the rule ‍Ij ! Ik ‍is an association rule if and
only if Conf(I
‍ j ! Ik )  min_Conf ‍and ‍Lift(Ij ! Ik )  min_Lift .‍
In association rule ‍Ij ) Ik ,‍ ‍Ij ‍is called antecedent or left-­hand-­side and ‍Ik ‍conse-
quent or right-­hand-­side, and it can be interpreted by “if ‍Ij ‍then ‍Ik ‍.” The implication
of this association rule is that if items contained by ‍Ij ‍ occur, there is a high (which
is guaranteed by ‍min_Conf )‍ and also higher (which is guaranteed by ‍min_Lift )‍
 Association rule learning  131

probability that the items in ‍Ik ‍ can be detected. The thresholds of Confidence
‍ ‍ and
‍Lift ‍show the basic idea of association rule generation: the threshold of Confidence
‍ ‍
guarantees that the ratio of ‍Sup(Ij \ Ik )‍and ‍Sup(Ij )‍is above a certain level, i.e., the
rule is meaningful, and the threshold of ‍Lift ‍ guarantees that the association rule is
strong enough to be regarded as an effective “rule.” In the following sections, the
most popular algorithm to generate association rules, namely, Apriori, will be intro-
duced in detail first. Then, as an improvement of Apriori, FP-­growth, which uses a
more efficient data structure to store the data and realize more efficient searching, is
then briefly covered.

10.2 Apriori algorithm

Before presenting the details of the Apriori algorithm, we first present the following
two properties of large item sets:
Property 1: Any non-­empty and strict subset of a large item set is large.
Property 2: Any super-­set of a non-­large item set cannot be large.
The above properties are intuitive. For Property 1, suppose we have a large item
set ‍ I‍, and it has two mutually exclusive and non-­empty item sets ‍Ij ‍ and ‍Ik ‍ where
 
‍Ij [ Ik = I .‍ The number of times where ‍ I ‍occurs in the database should be no less
than ‍min_Sup‍because it is a large item set. As the individual occurrence time of ‍Ij ‍
in the database should be no less than the occurrence time of ‍ I‍in the database, ‍Ij ‍is
also a large item set. This also applies to item set ‍Ik ‍. For Property 2, for a non-­large
item set ‍ I0 ‍, its occurrence time in the database is less than ‍min_Sup,‍ and thus the
occurrence time of any super-­set of ‍ I0 ‍is also less than ‍min_Sup.‍ Therefore, it cannot
be a large item set.
The Apriori algorithm is proposed by Agrawal et al. in 1993 [1]. It is devel-
oped based on the above properties to improve the computational efficiency of
large item set generation. Denote a large item set containing ‍n0 ‍ items as a large
‍n0 ‍-­item set. Denote ‍Ln0 ‍ as the set of all large ‍n0 ‍-­item sets. Denote ‍N(In0 )‍ as the
occurrence times of item set ‍In0 ‍ (‍In0 2 Ln0 )‍ in database ‍ D‍ . The algorithm to
generate the set of all large ‍n0 ‍-­ item sets, i.e., ‍Ln0 ,‍ is shown in Algorithm 1.

‍ ‍
132  Machine learning and data analytics for maritime studies

‍ ‍

In Algorithm 1, Step 1 aims to find all large 1-­item sets by scanning the whole
database ‍ D‍and examining each item it contains. When ‍n0  2‍, by iteration, candi-
date large ‍n0 ‍-­item sets are first found by combining the large item sets in ‍Ln0 1‍where
the first ‍(n0  2)‍items are the same while the last item is different, assuming that the
items are ordered in alphabet. Then, subsets containing ‍(n0  1)‍items of each com-
bination are checked to ensure that all of these subsets are large item sets. Otherwise,
the item set containing ‍n0 ‍ items should be removed from the set of candidate large
item sets. The detailed steps are shown in Algorithm 2. Then, ‍Support ‍of each of the
candidate large ‍n0 ‍-­item sets generated by Algorithm 2 is checked, and the sets with
0
‍Support ‍no less than ‍min_Sup‍are finally retained as the large ‍n ‍-­item sets, i.e., ‍Ln0 ‍.

‍ ‍
 Association rule learning  133

The following toy example is used to show the procedure of using Apriori
algorithm to generate large item sets from a given database ‍ D‍. We require that
‍min_Sup = 0.5‍. The set of records is shown in Table 10.1.
The candidate large 1-­item sets and final large 1-­item sets can be generated
by Step 1 of Algorithm 1. To be more specific, the candidate large 1-­item sets are
shown in Table 10.2, and the large 1-­item sets are shown in Table 10.3.
Then, all the large n'-­item sets, where n'>=2, are generated by Step 2 of
Algorithm 1, which combines the identified large (n'-1)-­item sets one by one if they
have the same first (n'- 2) items. Then, subsets of each candidate large item set are
evaluated. The candidate large 2-­item sets and the final large 2-­item sets are shown
in Tables 10.4 and 10.5, respectively.
Based on the large 2-­item sets generated, only one candidate large 3-­item set
f2,
‍ 3, 5g‍ can be formulated, which is to combine f2, ‍ 3g‍ and f2,
‍ 5g‍ as they have the
same first item 2. The ‍Support ‍ of f2,
‍ 3, 5g‍ is ‍
2 /4 ‍ and is equal to the threshold, and

Table 10.1   Records in the database

Records Items
t1
‍ 3, 4g‍
f1,
t2
‍ 3, 5g‍
f2,
t3
‍ 2, 3, 5g‍
f1,
t4
‍ 5g‍
f2,
134  Machine learning and data analytics for maritime studies

Table 10.2   The candidate large 1-­item sets

Candidate large 1-­item set ‍Support ‍

f1g
‍ ‍ ‍2/4‍
f2g
‍ ‍ ‍3/4‍
f3g
‍ ‍ ‍3/4‍
f4g
‍ ‍ 1‍ /4‍
f5g
‍ ‍ ‍3/4‍

thus it is a large 3-­item set. The process of large item set generation terminates, as
no more large item sets can be formulated.
After generating all large ‍n0 ‍-­ item sets, association rules can be gener-
ated from them. Thresholds for Confidence
‍ ‍ and ‍Lift ‍ denoted by ‍min_Conf ‍ and
min_Lift
‍ ,
‍ respectively, should be specified before association rule generating.
In Apriori algorithm, Property 3 is used to guide the generation of association
rules:
Property 3: Partitioning a large item set ‍I‍into two mutually exclusive and
complementary subsets ‍Ij ‍ and ‍Ik ‍. The rule from ‍Ij ‍ to ‍Ik ‍ is denoted by ‍Ij ! Ik .‍
Assume that Conf(I
‍ j ! Ik )_Conf .‍ Then, for any non-­empty and strict subset of ‍Ij ,‍
denoted by ‍Ij0 ,‍ and the superset of ‍Ik ,‍ denoted by ‍I0k = Ik  Ij0 ,‍ the rule from ‍Ij0 ‍to ‍I0k ‍
is called a sub-­rule of ‍Ij ! Ik ,‍ and Conf(I 0 0
‍ j ! Ik )_Conf .‍ The proof of Property 3
is as follows.
Proof: Denote the events of observing ‍Ij ‍ and ‍Ik ‍ by ‍E(Ij )‍ and ‍E(Ik )‍, respec-
tively, and the events of observing ‍Ij0 ‍ and ‍I0k ‍ by ‍E(Ij0 )‍ and ‍E(I0k ),‍ respectively. The
Confidence
‍ ‍of rule ‍Ij ! Ik ‍can be calculated by
P(E(Ij ) \ E(Ik )) P(E(I ))
Conf(Ij ! Ik ) = = , (10.3)
‍ P(E(Ij )) P(E(Ij )) ‍

Table 10.3   The large 1-­item sets

Large 1-­item set ‍Support ‍

f1g
‍ ‍ ‍2/4‍
f2g
‍ ‍ ‍3/4‍
f3g
‍ ‍ ‍3/4‍
f5g
‍ ‍ ‍3/4‍
 Association rule learning  135

Table 10.4   The candidate large 2-­item sets

Candidate large 2-­item set ‍Support ‍

‍ 2g‍
f1, 1‍ /4‍
‍ 3g‍
f1, ‍2/4‍
‍ 5g‍
f1, 1‍ /4‍
‍ 3g‍
f2, ‍2/4‍
‍ 5g‍
f2, ‍3/4‍
‍ 5g‍
f3, ‍2/4‍

and the Confidence 0 0

‍ ‍of rule ‍Ij ! Ik ‍can be calculated by
P(E(Ij0 ) \ E(I0k )) P(E(I ))
Conf(Ij0 ! I0k ) = = . (10.4)
‍ P(E(I ))
0
j P(E(Ij0 )) ‍

As ‍Ij0  Ij ,‍ we have ‍P(E(Ij0 ))(E(Ij ))‍ and thus Conf(I

‍0
0 0
j ! Ik )(Ij ! Ik ) < min_Conf .‍
Therefore, it can be concluded that Conf(I ‍ j ! Ik ) < min_Conf. ‍‍‍
0

Property 3 shows that a sub-­rule of a rule whose Confidence ‍ ‍ is less than

‍min_Conf ‍ has a smaller Confidence
‍ ‍ value, and thus, it cannot be an association
rule. This property is used in Apriori algorithm to simplify the process of generating
all association rules from a large item set by not examining sub-­rules of a rule with
0 
Confidence
‍ ‍ less than ‍min_Conf .‍ Given a large ‍n ‍-­item set ‍In0 ,‍ a consequent with
0 
‍m (1  m )‍ items is denoted by hm and the set of all hm derived from ‍In0 ‍ is denoted
by ‍Hm‍. In Apriori, all rules with Confidence
‍ ‍ no smaller than ‍min_Conf ‍ are first
generated, and then a rule is deleted if its ‍Lift ‍is less than ‍min_Lift .‍ The Algorithm
of association rule generation is presented in Algorithm 3.
Especially, Line 10 of Algorithm 4 is based on Property 3, which deletes the
rules with Confidence
‍ ‍less than ‍min_Conf ‍and thus all the sub-­rules of this rule will
not be considered, so as to reduce the computational burden. Line 14 of Algorithm 4
calls Algorithm 2 to generate ‍Hm‍from all pairs of item sets in ‍Hm1‍.

Table 10.5   The large 2-­item sets

Large 2-­item set ‍Support ‍

‍ 3g‍
f1, ‍2/4‍
‍ 3g‍
f2, ‍2/4‍
‍ 5g‍
f2, ‍3/4‍
‍ 5g‍
f3, ‍2/4‍
136  Machine learning and data analytics for maritime studies

Figure 10.1   Generation of association rules from large 3-­item set f2,
‍ 3, 5g‍

The procedure of generating association rules from the large 3-­item set f2, ‍ 3, 5g‍
found in the toy example is shown in Figure 10.1. Suppose that we set ‍min_Conf = 1‍
and ‍min_Lift = 1.2‍.
As shown in Figure 10.1, in Step 1, three candidate rules can be generated:
‍2, 3 ! 5‍, ‍2, 5 ! 3‍, and ‍3, 5 ! 2‍. Their Confidence
‍ ‍ values are first calculated, and
it is found that Confidence
‍ ‍of rule ‍2, 5 ! 3‍is less than ‍min_Conf = 1‍. Hence, its ‍Lift ‍
value is not calculated, and its sub-­rules are not considered, which are shown by the
gray nodes. Then, the ‍Lift ‍values of the other two rules are checked, which are larger
than the threshold of ‍Lift ‍, and thus these two rules are association rules. Then, in Step
2, only one candidate rule needs to be further investigated, namely, ‍3 ! 2, 5‍. This is
because the other two candidate association rules: ‍2 ! 3, 5‍and ‍5 ! 2, 3‍are sub-­rules
of rule ‍2, 5 ! 3‍ whose Confidence
‍ ‍ is smaller than ‍min_Conf = 1‍. As ‍3 ! 2, 5‍ has
Confidence
‍ = 2/3‍, which is less than ‍min_Conf = 1‍, it cannot be an association rule.
To conclude, two association rules can be generated from the large 3-­item set f2, ‍ 3, 5g‍
in the toy example: ‍2, 3 ! 5‍and ‍3, 5 ! 2‍.

‍ ‍
 Association rule learning  137

‍ ‍

Association rule learning is a popular approach to analyzing the relations

between deficiency codes as well as ship characteristics and the inspection results
including deficiencies and detention in ship PSC inspection in the existing literature.
Example 10.1 is to mine the relations between deficiency codes in ship inspection
records using a sub-­dataset of the whole inspection data set with 200 inspection
records. Weka, which is an open source software providing various tools for data
preprocessing, data mining and machine learning algorithms implementation, and
data visualization developed by researchers from The University of Waikato [2], is
used as the tool to mine the association rules regarding the deficiency codes from
the 200 inspection records.
Example 10.1: According to the Tokyo MoU [3], there are a total of 17 defi-
ciency codes, which are the inspection targets of a PSC inspection conducted by
each member state of the Tokyo MoU as listed in Table 10.6. Our aim is to figure
out the correlations between the deficiency codes except for D99-­Other, as this code
does not have a specific meaning. Especially, the correlation is presented by the
large item sets found and the association rules mined from them.
The deficiency information in the initial data collected online is specific defi-
ciency items, e.g., “01103” representing “Passenger Ship Safety (including exemp-
tion),” “07101” representing “Fire prevention structural integrity,” and “12101”
representing “Stowage/segregation/packaging dangerous goods.” Therefore, the
initial data are first pre-­processed by generating 16 features with each feature repre-
senting a deficiency code, i.e., D01, D02, …, D15, D18, and the value of each record
for each feature is the number of deficiencies of the record under this deficiency
code. To enable the Weka software to generate association rules from the data set
138  Machine learning and data analytics for maritime studies

Table 10.6   Deficiency codes in the Tokyo MoU [3]

Deficiency Content Deficiency code Content

code
D01 Certificate and D10 Safety of navigation
documentation
D02 Structural conditions D11 Life-­saving appliances
D03 Water/weathertight D12 Dangerous goods
conditions
D04 Emergency systems D13 Propulsion and auxiliary
machinery
D05 Radio communications D14 Pollution prevention
D06 Cargo operations including D15 International safety
equipment management
D07 Fire safety D18 Labor Conditions
D08 Alarms D99 Other
D09 Working and living
conditions

containing deficiency codes of each record, all non-­zero feature values in the data
set are encoded to “T”, and zero feature values are encoded to “?”. Then, the initial
.csv file is converted to .arff file. The thresholds of ‍Support ‍, Confidence
‍ ‍, and ‍Lift ‍
are set to ‍min_Sup = 0.1‍, ‍min_Conf = 0.7‍, and ‍min_Lift = 1.5‍. The large item sets
are shown in Tables 10.7–10.9, and the association rules mined from the large item
sets are shown in Table 10.10.
Table 10.7 shows that the most frequently detected deficiency items are D07:
fire safety, which is identified in more than half of the 200 inspections of concern,
followed by D10: safety of navigation, which is found in 43.5% of the inspections.
From Table 10.8, it can be seen that deficiency codes D07 and D10 are often detected
together, which can be found in nearly 25% of the inspection records. In addition,
D04 and D07 as well as D07 and D11 are often identified in one inspection, which

Table 10.7   Large 1-­item sets with ‍min_Sup = 0.1‍

Large item set Frequency (support)

D07 102(0.510)
D10 87(0.435)
D04 50(0.250)
D18 48(0.240)
D14 46(0.230)
D11 43(0.215)
D03 41(0.205)
D13 29(0.145)
D09 28(0.140)
 Association rule learning  139

Table 10.8   Large 2-­item sets with ‍min_Sup = 0.1‍

Large item set Frequency (‍Support )‍

D07, D10 48(0.240)
D04, D07 32(0.160)
D07, D11 30(0.150)
D07, D18 28(0.140)
D07, D14 26(0.130)
D10, D11 26(0.130)
D03, D07 25(0.125)
D03, D10 25(0.125)
D04, D10 25(0.125)
D10, D14 24(0.120)
D04, D11 21(0.105)
D10, D18 21(0.105)

can be seen in more than 15% of the inspection records. Finally, Table 10.9 shows
that D07, D10, and D11 together with D04, D07, and D10 are detected together in
more than 10% of the inspection records. The association rules shown in Table 10.10
are all generated from the large 3-­item sets and can be interpreted as follows. For
example, for the association rule D10, D11 ) ‍ ‍D07, if D10 and D11 are detected on
one ship, the probability that the ship has deficiency code D07 is 81%. Compared to
the fact that the probability that a ship has deficiency D07 is 51% if no prior informa-
tion is known as shown in Table 10.7, the ‍Lift ‍value of this association rule is about
1.58, showing that identifying deficiency codes D10 and D11 can increase the prob-
ability of identifying deficiency code D07 by 1.58 times. Therefore, this association
rule can be used to guide onboard ship inspection: if deficiency codes D10 and D11
are detected on a ship, it is suggested that the PSC inspector should pay more atten-
tion to deficiency items under deficiency code D07 as there is a high probability that
the ship has a deficiency/deficiencies under this code.

10.3 FP-growth algorithm

The main difference between the FP-­growth algorithm and the Apriori algorithm is
in the context of generating all large item sets, that is, finding all item sets with

Table 10.9   Large 3-­item sets with ‍min_Sup = 0.1‍

Large item set Frequency (‍Support ‍)

D07, D10, D11 21(0.105)
D04, D07, D10 20(0.100)
140  Machine learning and data analytics for maritime studies

Table 10.10   Association rules mined with ‍min_Conf = 0.7‍and ‍min_Lift = 1.5‍

Association rule Confidence

‍ ‍ ‍Lift ‍
D10, D11 )
‍ ‍D07 0.81 1.58
‍ ‍D07
D04, D10 ) 0.80 1.57
‍ ‍D10
D07, D11 ) 0.70 1.61

‍Sup(In0 )  min_Sup‍, while the procedure of generating all association rules is the
same in both algorithms. FP-­growth algorithm is proposed by Han et al. in 2000 [4]
and is believed to be an improvement of Apriori as it is much faster. In Apriori, data
are stored in set structure, and candidate large item sets should be generated, the num-
ber of which can be large. Moreover, the whole database needs to be traversed multi-
ple times to check the ‍Support ‍values of the candidate item sets in a brute-­force man-
ner, and traversing the whole data set heavily reduces the computational efficiency. In
the FP-­growth algorithm, data are stored in a tree structure called a frequent pattern
tree or FP tree. FP tree allows for faster scanning, while no candidate large item sets
are to be generated. In an FP tree, each node represents a single item in the original
database, and the links between the nodes represent their co-­occurrence in the data-
base. We do not show the detailed steps to construct an FP tree and the large item sets
in this section, and we only present the rough steps as follows. Readers are referred
to Han et al. (2001) [4] and Borgelt [5] for more details.

1. Scan the database to calculate the frequency (i.e., ‍Support )‍ of individual items,
and remove the items with ‍Support ‍values less than ‍min_Sup.‍
2. Start to construct the FP tree by creating the root of the tree that represents null.
3. Scan the database again to exam each of the records and add them to the tree one
by one. Especially, the branch of the tree is constructed with items in descend-
ing order of count.
4. After constructing the FP tree, the tree is mined from the lowest nodes as well as
the links to the nodes. The lowest nodes represent the frequency pattern length
1. From each of the lowest nodes, traverse the path in the FP Tree, and the path
or paths are called a conditional pattern base.
5. Construct a conditional FP tree based on the conditional pattern bases of the
nodes.
6. Large item sets are generated from the conditional FP tree constructed.

References

[1] Agrawal R., Imieliński T., Swami A. ‘Mining association rules between sets of
items in large databases’. The 1993 ACM SIGMOD International Conference
 Association rule learning  141

[online]; Washington, DC, New York, NY, 1993. pp. 207–16. Available from
http://portal.acm.org/citation.cfm?doid=170035
[2] Witten I.H., Frank E. ‘Data mining’. ACM SIGMOD Record. 2002;31(1):76–77.
[3] The Tokyo MOU; 2017 List of tokyo mou deficiency codes [online]]]. Available
from http://www.tokyo-mou.org/doc/Tokyo%20MOU%20deficiency%​20codes
%20(December%202017).pdf [Accessed 9 Aug 2021].
[4] Han J., Pei J., Yin Y. ‘Mining frequent patterns without candidate generation’.
ACM SIGMOD Record. 2000;29(2):1–12.
[5] Borgelt C. An implementation of the FP-­growth algorithm [online]. 2005:1–5.
Available from http://portal.acm.org/citation.cfm?doid=1133905
This page intentionally left blank
 Chapter 11
Cluster analysis

Cluster analysis, or clustering, is a general task aiming to group a given set of examples
into several groups (i.e., clusters) following given criteria, such that the examples in
the same group are as close as to each other, and the examples in different groups are as
different as from each other. Generally, clustering works in the context of unsupervised
learning, where only the features of the examples are known while there is no target
defined or targets are unknown. It aims to divide the whole data set into several mutu-
ally exclusive and complementary clusters, so as to mine the properties of the clusters
formulated where such properties are represented by cluster labels. Mathematically,
suppose that we have an unlabeled data set ‍D = fxi , i = 1, ..., ng‍, where ‍xi 2 Rm‍ is
the feature vector that can be represented by ‍xi = (xi1 , ..., xim )‍. A clustering algorithm
aims to divide ‍ D‍ into ‍ K ‍ groups fC 0 00
‍ k |k = 1, ..., Kg‍ such that C
‍ k0 \ Ck00 = ;, k ¤ k ‍,
0 00 K
‍k 2 f1, ..., Kg‍ and ‍k 2 f1, ..., Kg‍, and ‍D = [k=1 Ck .‍ The cluster label of cluster C ‍ k‍
is denoted by ‍k ‍, which needs to be decided and summarized by model user. Based
on the cluster labels, the label of each example, which we call it by example cluster
label, can be obtained and is denoted by ‍0i ‍, e.g., ‍xi‍. One needs to distinguish between
clustering and classification, although both aim to separate examples in the data set. In
classification, labels of the examples are known, and the model training process aims
to separate the data according to the features while minimizing a given loss function
measured by the examples’ labels, so as to obtain a classifier to predict the target of the
new examples. In contrast, labels are unavailable in clustering, and the model training
process aims to put the examples similar to each other regarding their features in the
same cluster, and then to summarize the cluster labels of the clusters generated.
In the following sections, distance measures of two examples in the data set and
of two clusters are first introduced, as distance measure is the most important con-
cept in cluster analysis. Then, metrics to evaluate the performance of cluster analysis
are introduced. Finally, popular clustering algorithms are discussed with examples
of ship inspection by PSC provided.

11.1 Distance measure in clustering

11.1.1 Distance measure of examples

The key point in cluster analysis is how to measure the “similarity” between two
examples in the data set, and this is usually achieved by the calculation of “distance”
144  Machine learning and data analytics for maritime studies

of these two examples. Distance measure is an objective score used to measure the
relative difference/dissimilarity between two examples in the problem of concern.
The distance between two examples ‍xi ‍and ‍xj ‍is denoted by ‍dist(xi , xj ),‍ which satis-
fies the following properties:

1. non-­negativity: ‍dist(xi , xj )  0;‍

2. identity: If and only if ‍xi = xj ,‍ ‍dist(xi , xj ) = 0;‍
3. symmetry: ‍dist(xi , xj ) =dist(xj , xi );‍ and
4. triangle inequality: ‍dist(xi , xj )  dist(xi , xk ) + dist(xk , xj ).‍

Features of an example can be numerical and categorical. Numerical features

are ordinal, where the relative feature values are comparable. For example, ship age
is a numerical feature, where a ship of age 5 is younger than a ship of age 10 by 5
years. Categorical features can be either ordinal, where the relative feature values
are comparable like numerical features (e.g., low, medium, and high for ship com-
pany performance, where a ship company with high performance is better than a
ship company with medium performance, and is much better than a ship company
with low performance), or nominal, where the feature values only indicate the cat-
egories and cannot be compared (e.g., container ship, bulk carrier, and passenger
ship belonging to the feature of ship type, and they cannot be compared directly
with each other). As feature values are comparable in ordinal features and noncom-
12 Applications
parable in nominalof features,
machine different
learning and dataofanalytics
means distance models
measure should be used
in these two types of features. For data set D with ‍m‍ features, denote the number
of its ordinal
Table 10.10 features
Association and mined
by m1rules the number of its nominal
with min_Conf = 0.7 features by m2=, 1.5
and min_Lift where
m
‍ = m 1 + m .
2‍ For ordinal features, Minkowski distance taking the following form is
the most popular one:
Association rule Confidence Lift
Pm1 ˇ ˇp  1p
D10, dist
D11 ) (x
D07 , x ) = ˇxim0.81
0  xjm0 ˇ ,‍ 1.58 (11.1)
‍ mk i j m 0 =1
D04, D10 ) D07 0.80 1.57
D07, D11 ) D10 0.70 1.61
where the subscript is ‍mk ‍short for Minkowski, and p should be no less than 1, such
that the properties of distance measure can be satisfied. Common values of p are 1
and 2. When Pm p = 1, Equation (11.1) is also called Manhattan distance, and can be
‍ 0 ‍0
written as m0 =1 wm k xm , k = 1, : : : , K.
tk =

  Pm1 ˇ ˇ
distman (xi , xj ) =xi  xj 1 = ˇxim0  xjm0 ˇ . (11.2)
‍ m0 =1 ‍
Manhattan P
m1
distance
= xm0 A0 xm0 B0is. the sum of the absolute difference between each pair of com-
ponents mof0 =1 the coordinates of two examples whose calculation is related to the L1

vector norm. It was first proposed to calculate the distance of driving from one
crossroad to another crossroad in Manhattan. Therefore, Equation (11.2) is also
10.3
referred FP-growth
to as city blockalgorithmdistance.
When ‍p = 2,‍ Equation (11.1) is also called Euclidean distance, and can be writ-
The
ten asmain difference between the FP-growth algorithm and the Apriori algorithm is
in the context of generating all large item sets, that is, finding all item sets with
Sup(In0 )  min_Sup, while the procedure of generating all association rules is the
same in both algorithms. FP-growth algorithm is proposed by Han et al. in 2000 [4]
and is believed to be an improvement of Apriori as it is much faster. In Apriori, data
are stored in set structure, and candidate large item sets should be generated, the num-
ber of which can be large. Moreover, the whole database needs to be traversed multi-
ple times to check the Support values of the candidate item sets in a brute-force man-
 Cluster analysis  145
s
Pm1 ˇ ˇ
disted (xi , xj ) = ||xi  xj ||2 = ˇxim0  xjm0 ˇ2 . (11.3)
m0 =1
‍ ‍
Euclidean distance is the most widely used distance measure of two examples whose
calculation is related to the L2 vector norm, and can be interpreted as the linear dis-
tance between two examples.
To measure the similarity of two examples regarding the angle between them,
one can use cosine similarity, which is calculated as follows:
Pm1
xi  xj 0 xim0 xjm0
cos(xi , xj ) = = qP m =1 qP . (11.4)
|xi |  |xj | m1
x 2 m1
x2
m0 =1 im0 m0 =1 jm0
‍ ‍
The range of the cosine similarity of two examples is [‍ 1, 1]‍. Then, the cosine dis-
tance of two examples can be calculated from the cosine similarity of them by
‍ distcos (xi , xj ) = 1  cos(xi , xj ).‍ (11.5)
One should also note that, for two vectors that are normalized by L2 normalization,
the squared Euclidean norm of their difference is equivalent to twice of their cosine
distance, whose proof is as follows:
Proof: For two vectors ‍A = (x1A , x2A , ..., xm1 A )‍and ‍B = (x1B , x2B , ..., xm1 B ),‍ they
are first normalized by L2 normalization, respectively, and we have
x1A x2A xm1 A
A0 = ( q ,q , ..., q ) = (x1A0 , x2A0 , ..., xm1 A0 )
x21A + x22A + ... + x2m1 A x21A + x22A + ... + x2m1 A x21A + x22A + ... + x2m1 A
‍ ‍
and
x1B x2B xm1 B
B0 = ( q ,q , ..., q ) = (x1B0 , x2B0 , ..., xm1 B0 )
x21B + x22B + ... + x2m1 B x21B + x22B + ... + x2m1 B x21B + x22B + ... + x2m1 B
‍ ‍
Then, the cosine distance between ‍ A ‍ and ‍ B ‍can be calculated by
0 0

Pm1
0 xm0 A0 xm0 B0
cos(A0 , B0 ) = qP m =1 qP
m1 m1
x2
m0 =1 m0 A0
x2
m0 =1 m0 B0
(11.6)
Xm1

= xm0 A0 xm0 B0 .
‍ m0 =1 ‍
The squared Euclidean norm of the difference of ‍ A0 ‍ and ‍ B0 ‍has the following form
|A0  B0 |2 = (x1A0  x1B0 )2 + (x2A0  x2B0 )2 + ... + (xm1 A0  xm1 B0 )2
P
m1
= 22 xm0 A0 xm0 B0 . (11.7)
m0 =1

‍ = 2  2 cos(A0 , B0 ) ‍
Then, the equivalence of the squared Euclidean norm and twice of the cosine dis-
tance of two vectors is proven.‍‍
Modified cosine similarity aims to reduce the influence of the absolute values
of the components of a feature vector. It is calculated by subtracting the mean of
146  Machine learning and data analytics for maritime studies

components of the vector for an example from each component of the vector. To be
more specific, the mean of ‍xi ‍(which is of m1 dimensions) is calculated by
Pm1
0 xim0
xi = m =1 , (11.8)
‍ m1 ‍
and the mean of ‍xj ‍is calculated by
Pm1
0 xjm0
xj = m =1 . (11.9)
‍ m1 ‍
Modified cosine similarity can be calculated by
(xi  xi )  (xj  xj )
cos0 (xi , xj ) =
|xi  Pxi |  |xj  xj |
m1
m0 =1
(xim0  xi )(xjm0  xj ) . (11.10)
= pP qP
m1 m1
m0 =1
(xim0  xi )2 m0 =1
(xjm0  xj )2
‍ ‍
Similarly, the revised cosine distance can be calculated by
‍ dist0cos (xi , xj ) = 1  cos0 (xi , xj ).‍ (11.11)
Measuring distance of examples containing nominal features can be much more
difficult, as their distance can hardly be measured because feature values are
noncomparable. To measure the distance of two examples containing nominal
features, cosine distance shown in Equation (11.5) and modified cosine distance
shown in Equation (11.11) can be used, as they only consider the angle of the
feature vector instead of the absolute values. If the nominal features are encoded
into binary features (e.g., the three values for feature ship type, namely, con-
tainer ship, bulk carrier, and passenger ship are encoded to three new features:
is_container_ship, is_bulk_carrier, and is_passenger_ship, and 100, 010, and
001 are used as the feature values of the three new features to indicate the three
types of ships), Hamming distance can be used. Hamming distance is a measure-
ment of the number of different values in two strings of equal length, and it can
be calculated by
Pm
‍ dist0ham (xi , xj ) = m20 =1 1(xim0 ¤ xjm0 ),‍ (11.12)

‍ ‍is the indicator function, which takes value 1 if ‍x ‍is true and 0, otherwise.
where 1(x)

11.1.2 Distance measure of clusters

In some cluster analysis methods (e.g., hierarchical clustering), distance between
clusters needs to be calculated. The distance between two clusters is usually cal-
culated based on the function of the pairwise distance between one example in one
cluster and one example in the other cluster. Given two clusters C ‍ k0 ‍ and C‍ k00 ‍ and a
distance measure of two examples ‍xi ‍ and ‍xj ‍ denoted by ‍dist(xi , xj )‍, their distance,
denoted by ‍Dist(Ck0 , Ck00 )‍, can be measured by
 Cluster analysis  147

1. Single-­linkage clustering:
‍Distmin (Ck0 , Ck00 ) = minxi 2Ck0 ,xj 2Ck00 dist(xi , xj ),‍ which is the minimum dis-
tance of two examples belonging to the two clusters, respectively.
2. Complete-­linkage clustering:
‍Distmax (Ck0 , Ck00 ) = maxxi 2Ck0 ,xj 2Ck00 dist(xi , xj ),‍ which is the maximum dis-
tance of two examples belonging to the two clusters, respectively.
3. Unweighted average-­linkage clustering: P
1
Distuavg (Ck0 , Ck00 ) = |C 0 ||C | xi 2Ck0 ,xj 2Ck00 dist(xi , xj ), which is the
‍ k k 00 ‍
average distance of all pairs of two examples belonging to the two clusters,
respectively.
4. Centroid-­linkage clustering:
‍Distcen (Ck0 , Ck00 ) =dist(ck0 , ck00 )‍ where ‍ck0 ‍ and ‍ck00 ‍ are the centroids of clus-
1 P
ters C
‍ k0 ‍ and C ‍, respectively, and can be calculated by ‍ck0 = |Ck0 | xi 2Ck0 xi ‍
‍ k00P
and ck00 = |C100 | xj 2C 00 xj . Therefore, centroid-­linkage clustering measures the
‍ k k ‍
distance between centroids of the two clusters.

11.2 Metrics for clustering algorithm performance evaluation

Metrics to evaluate the performance of supervised ML models (mainly referring to

regression and classification models) mainly aim to calculate the difference between the
predicted and actual target values, and smaller difference usually indicates a more accu-
rate prediction model, which is quite intuitive. Nevertheless, data sets used in cluster
analysis are not labeled, i.e., unsupervised. Therefore, metrics for supervised learning
are not applicable. Considering the fact that the core idea of clustering is that examples
in the same cluster should be close to each other, while examples in different clusters
should be far from each other, it can be seen that clusters of examples divided by a good
clustering model should have high intra-­cluster similarity while low inter-­cluster simi-
larity. Especially, two indicators are used to measure the distance between examples in
one cluster Ck, given a distance measure of two examples is denoted by
P
avg(Ck ) = |C |(|C2 |1) 1i|Ck | dist(xi , xj ), (11.13)
‍ k k ‍
which is the average distance of each pair of the examples belonging to cluster C
‍ k ,‍ and
‍ max(Ck ) = max1i|Ck | dist(xi , xj )‍, (11.14)
which is the largest distance of each pair of examples belonging to cluster C ‍ k .‍
One popular metric to evaluate the performance of a clustering algorithm is
called Davies–Bouldin index, or DBI for short, which considers average intra-­
cluster and inter-­cluster distances, is shown in Equation (11.15):
PK  
avg(Ck0 )+avg(Ck00 )
DBI = K1 k0 =1 maxk00 =1,...,k0 1,k0 +1,...,K Distcen ,, (11.15)
‍ (Ck0 ,Ck00 ) ‍
where a smaller value of the whole part in parentheses indicates better clustering
performance (as a smaller numerator shows smaller intra-­cluster distance and a
148  Machine learning and data analytics for maritime studies

larger denominator shows larger inter-­cluster distance). Then, the worst case of all
the clusters is identified by the ‍ max ‍ operator, and the DBI calculates the sum of
ratio between intra-­cluster distance and inter-­cluster distance in the worst case of all
the clusters generated. Therefore, the smaller the value of DBI, the better the per-
formance of a clustering model. Another popular metric is called Dunn index (DI),
which considers extreme intra-­cluster and inter-­cluster distances shown as follows:
min1k0 <k00 K Distmin (Ck0 , Ck00 )
DI = , (11.16)
‍ max1kK max(Ck ) ‍
where ‍max(Ck )‍ is calculated in Equation (11.14). A larger value of DI indicates
that a clustering algorithm performs better, as a larger minimum inter-­cluster dis-
tance (i.e., ‍Distmin (Ck0 , Ck00 ))‍ or a smaller maximum intra-­cluster distance (i.e.,
‍max1kK max(Ck )‍) increases the value of DI.
The above two metrics are called internal index for clustering performance
evaluating as they only consider the clusters generated by the clustering algorithm
itself. If there is a reference model that the clustering algorithm can be compared
with, external indexes can be used. For data set ‍ D‍ containing ‍n‍ examples, denote
the clusters given by the reference model by fC 
0
‍ l |l = 1, ..., Sg‍ with example cluster
‍ i |i = 1, ..., ng.‍ Recall that the clusters given by the current clustering model
label f
is fC 0
‍ k |k = 1, ..., Kg‍ with cluster label f
‍ i |i = 1, ..., ng.‍ Note that ‍S ‍ is not necessarily
equal to ‍ K ‍. Examples in ‍ D‍are divided into clusters using the reference model and
the current clustering model to be evaluated. For each two examples ‍xi 2 D‍ and
‍xj 2 D‍with i‍ ,‍ cases of how they belong to the clusters of the two clustering models
are as follows:
0 0
1. ‍SS = f(xi , xj )|0i = j0 , i  = j  , ig,‍ which means that ‍xi ‍ and ‍xj ‍ are divided into
the same cluster in both the clustering model to be evaluated and the reference
model. Denote the number of such pairs of examples by ‍a = |SS|;‍
0 0
2. ‍SD = f(xi , xj )|0i = j0 , i  ¤ j  , ig,‍ which means that ‍xi ‍and ‍xj ‍are divided into
the same cluster in the clustering model to be evaluated but different clusters in
the reference model. Denote the number of such pairs of examples by ‍b = |SD|;‍
0 0
3. ‍DS = f(xi , xj )|0i ¤ j0 , i  = j  , ig,‍ which means that ‍xi‍ and ‍xj ‍ are divided into
the same cluster in the reference model but different clusters in the clustering
model to be evaluated. Denote the number of such pairs of examples by ‍c = |DS|;‍
0 0
4. ‍DD = f(xi , xj )|0i ¤ j0 , i  ¤ j  , ig,‍ which means that ‍xi ‍ and ‍xj ‍ are divided
into different clusters in both the clustering model to be evaluated and the refer-
ence model. Denote the number of such pairs of examples by ‍d = |DD|.‍

As we require that i < j, each pair of examples only appears once in ‍SS ,‍ ‍SD,‍ ‍DS ,‍
and ‍ DD‍, and we have ‍a + b + c + d = n(n  1)/2.‍ Based on the above cases, three
external indexes, namely, Jaccard coefficient (JC), Fowlkes and Mallows index
(FMI), and rand index (RI), are defined as follows:
a
JC = , (11.17)
‍ a + b + c ‍
 Cluster analysis  149
q
a a
FMI = a+b  a+c ,‍ (11.18)
‍
and
2(a + d)
RI = . (11.19)
‍ n(n  1) ‍
The range of the above three indicators is [‍ 0, 1]‍, and larger values indicate better
clustering algorithm performance.

11.2.1 Clustering algorithms
Classic clustering algorithms can roughly be divided into three types: partition-­
based methods, density-­based methods, and hierarchy-­based methods. A summary
of these algorithms is presented in Table 11.1.
In the following subsections, a typical algorithm under each type of clustering
methods is introduced. In particular, we introduce K-­means as the representative of
partition-­based methods, density-­based spatial clustering of applications with noise
(DBSCAN) as the representative of density-­based methods, and agglomerative
algorithm as the representative of hierarchy-­based methods.

11.2.2 K-means (partition-based method)

K-­means is one of the simplest and the most widely used clustering algorithms.
Especially, K is the number of clusters to be generated and its value needs to be pre-
set, and “means” refers to averaging of the data by finding the centroids of each of
the K clusters. The basic step of K-­means algorithm is to find a total of ‍ K ‍ centroids
as the centers of the clusters and allocate the examples in the data set to them. The
positions of the ‍ K ‍centroids are constantly updated to improve the clustering perfor-
mance. The procedure of the K-­means algorithm is shown in Algorithm 1.

‍
Table 11.1   A summary of classic clustering algorithms

Type of clustering Basic idea Typical algorithms Advantages Disadvantages

methods
Partition-­based methods Regard the center of K-­means and K-­ In general, relatively The number of clusters (e.g., the value
examples belonging to a medoids low time complexity of ‍ K ‍in K-­means and K-­medoids)
cluster as the center of the and high computing should be preset, and the clustering
corresponding cluster. efficiency. results are highly sensitive to the
preset cluster number; highly
sensitive to outliers as they contribute
much to the distance calculated;
easily fall to local optimum; not
suitable for non-­convex data.
Density-­based methods Data in the region with high DBSCAN, OPTICS, High computational When the density of data space is
data space density are and mean-­shift. efficiency; suitable uneven, the resulting clusters might
considered to belong to to deal with data have low quality; when the data set is
the same cluster. set with arbitrary large, a big size memory is needed;
shape. highly sensitive to model parameters.
Hierarchy-­based methods Construct hierarchical Agglomerative Can deal with data set Relatively high time complexity; the
structure among examples algorithm and with arbitrary shape number of clusters needs to be preset.
for clustering. Each divisive algorithm. and feature types;
example in the data the hierarchical
set is first regarded as structure in the
a cluster, and the two data set is easily
closest clusters are merged detected.
to form a new cluster
until there is only one
cluster left. Or, a reversed
process.
 Cluster analysis  151

‍
It is noted that the stopping criterion of K-­means is that ‍ck ‍is not updated, which
means that the centroids of the clusters are not changed any more. This can take
quite a long time or lead to too many iterations. To reduce the computational burden,
the longest algorithm running time or the maximum number of updates can be set,
and the algorithm terminates when the threshold of the running time or the number
of updates is reached. The advantages of K-­means are that:

1. Relatively simple and easy to understand and implement;

2. Can guarantee convergence and is applicable to large-­scale data sets;
3. Can be easily generalized to clusters of different shapes and sizes, as well as
adapted to new examples.

Meanwhile, the disadvantages of K-­means are that:

1. The value of ‍ K ‍should be preset in a manual manner, and can be prone to biased
subjective judgment.
2. As the initial centroids are generated randomly, the algorithm suffers from
uncertainties.
3. As the centroids are determined by their distance to examples in the data set,
clustering performance can be heavily influenced by outliers.

11.2.3 DBSCAN (density-based method)

DBSCAN is short for density-­based spatial clustering of applications with noise,
whose main idea is to find core examples exhibiting high density and expand clus-
ters from them. In the meantime, outliers in low-­density region that do not belong
to any cluster can be marked. The basic idea of DBSCAN shows that there are two
main parameters in the DBSCAN algorithm:

1. ‍‍, which is a threshold of distance between two examples for one to be considered
as in the neighborhood of the other. For example, ‍xi 2 D,‍ its ‍‍-­neighborhood,
denoted by ‍N (xi ),‍ is constituted by the examples whose distance from ‍xi ‍is no
more than ‍‍, i.e., ‍N (xi ) =fxj 2 D|dist(xi , xj )  g.‍ This guarantees that exam-
ples in the same cluster are close enough to each other.
2. ‍Min_Pts,‍ which is the minimum number of examples in the neighborhood of a cer-
tain example (including the example itself) for this examples to be considered as a
core example. This guarantees that a cluster contains a certain number of examples.
152  Machine learning and data analytics for maritime studies

Clusters in DBSCAN are formed based on the following concepts:

1. Core example: for a given example ‍xi 2 D,‍ if the number of examples in its ‍
‍-­neighborhood is no less than ‍Min_Pts,‍ i.e., |N ‍  (xi )|  Min_Pts,‍ ‍xi ‍ is a core
example.
2. Directly density reachable: if example ‍xi ‍is a core example and ‍xj 2 N (xi ),‍ then
example ‍xj ‍ is directly density reachable from example ‍xi .‍ It should be men-
tioned that ‍xi ‍might not be directly density reachable from example ‍xj ,‍ unless ‍xj ‍
is a core example.
3. Density reachable: for examples ‍xi ‍and ‍xj ,‍ there is a series of examples ‍xk1 , ..., xkl ‍
such that ‍xk1 = xi ‍and ‍xj ‍is directly density reachable from ‍xkl .‍ If ‍xkl0 ‍and ‍xkl0 +1,‍
where 1‍  l0 ,‍ are directly density reachable, meaning that ‍xkl0 ‍ is within the ‍
‍-­neighborhood of ‍xkl0 +1,‍ and ‍xkl0 +1‍ is also within the ‍‍-­neighborhood of ‍xkl0 ,‍
then ‍xj ‍is density reachable from ‍xi .‍ As the series of examples ‍xk1 (i.e., xi ), ..., xkl ‍
are directly density reachable to each other, they must be core examples.
Meanwhile, ‍xj ‍is not necessary a core example.
4. Density connected: for examples ‍xi ‍ and ‍xj ,‍ if there is another example ‍xk ‍ such
that ‍xi ‍is density reachable from ‍xk ,‍ and ‍xj ‍is density reachable from ‍xk ,‍ then ‍xi ‍
and ‍xj ‍are density connected. For these three examples ‍xi ,‍ ‍xj ,‍ and ‍xk ,‍ ‍xk ‍must be
a core example, while ‍xi ‍and ‍xj ‍are not necessarily core examples.

Figure 11.1 is used to further illustrate the above concepts. Given the condition
that ‍Min_Pts = 3‍, examples in red are core examples as no less than three examples
are in their ‍‍-­neighborhood, while those not in red are noncore examples. Examples
that are directly density reachable from a core example are in the blue circle with
the core example in red at the center. The green links connecting the core exam-
ples show that these core examples are density reachable with each other. While

Figure 11.1   An example to illustrate the concepts in DBSCAN algorithm

 Cluster analysis  153

if a core example is within the ‍‍-­neighborhood of another core example, they are
directly reachable to each other. As examples in blue are within the ‍‍-­neighborhood
of the core examples that are density reachable, all the examples in blue are density
connected.
Based on the above concepts and explanation, a cluster in DBSCAN is defined
as a maximum set of examples that are density connected from a core example.
In other words, for a core example ‍xi ,‍ a cluster expanded from it is denoted by

‍ k = fx 2 D|x and xi are density connectedg.‍ The procedure of the DBSCAN algo-
C
rithm is shown in Algorithm 2.
From lines 2 to 7 of Algorithm 2, the set of all core examples is first gener-
ated by traversing the whole data set. Then, a random core example is selected,
and all the core examples that are density reachable from this core example are
found. Especially, to ensure that the cluster is expanded from the initial random core
example, queue data structure (i.e.,  ‍ cur ‍) is used to store the core examples density
reachable from this core example, so as to ensure that these core examples are vis-
ited one by one from the nearest to the farthest from the initial core example. Then,
examples that are in the ‍‍-­neighborhood of all the core examples are added to the
current cluster. This is shown from lines 14 to 18. Therefore, the initial (starting)
core example is density connected with the other examples in this cluster. The above
process is repeated until all the core examples are processed (i.e., added to clusters),
which is shown by the while loop from lines 10 to 22 of Algorithm 2.

‍
154  Machine learning and data analytics for maritime studies

‍
Unlike K-­means where all the examples are divided into clusters, there might be
some examples that do not belong to any cluster in DBSCAN, which are referred to
as noises. The advantages of DBSCAN are that:
1. It is nonsensitive to noises and can also distinguish noise examples from normal
examples in the data set.
2. The number of clusters does not need to be preset.
3. It can generate clusters of arbitrary shapes, which means that the contour of the
examples contained by a cluster can be of any shape, as long as the examples in
one cluster are density connected from the initial core example.

Meanwhile, the disadvantages of DBSCAN are that:

1. It is ineffective when dealing with data sets with uneven density, which means
that the distribution of the examples in the data set is uneven in the feature
space.
2. When the size of the data set is large, DBSCAN would take a long time to
converge.
3. The values of parameters ‍‍ and ‍Min_Pts‍can have a very large influence on the
clusters generated.
4. As a random unvisited core example is selected as the seed from which a cluster
is to be generated, the stability of the algorithm is adversely influenced.

11.2.4 Agglomerative algorithm (hierarchy-based methods)

Hierarchy-­based methods can be either bottom-­up (i.e., agglomerative clustering)
or top-­down (i.e., divisive clustering), where examples in the former are viewed
as individual clusters at the beginning, and then merged with each other based on
cluster distance until the preset number of clusters is reached, while examples in
the latter are in one cluster at the beginning, and are gradually divided into sub-­
clusters based on cluster distance until the preset number of clusters is reached.
The calculation of cluster distance is introduced in section 11.1.2, and the first three
 Cluster analysis  155

cluster distance calculation methods, namely, single-­linkage clustering, complete-­

linkage clustering, and unweighted average-­linkage clustering, can be used for clus-
ter distance measuring in hierarchy-­based methods. In this section, we introduce the
agglomerative algorithm using unweighted average-­linkage clustering to measure
cluster distance as the representative of hierarchy-­based methods. The basic idea of
agglomerative algorithm is intuitive. Each example in the data set is viewed as an
individual cluster at the beginning. Then, in each round of iteration, it constantly
merges two clusters with the smallest cluster distance, until the number of clusters
generated reaches the preset number of clusters. The procedure of agglomerative
algorithm is presented by Algorithm 3.
From lines 1 to 5 of Algorithm 3, the clusters as well as the cluster distance
matrix are initialized, where each example in the data set is regarded as an individual
cluster. Then, from lines 6 to 15, a pair of clusters with the smallest cluster distance
are merged into one cluster one by one in each iteration, and the distance matrix is
also updated accordingly, until the number of preset clusters is reached, and then the
algorithm terminates.

‍ ‍

We use Example 1 to show how to use cluster analysis to cluster ships in PSC
inspection while generating explanations and implications from the results using
156  Machine learning and data analytics for maritime studies

K-­means as an example. To show the clustering process visually, we use a data set
containing 50 ships that are randomly chosen from the whole data set with two fea-
tures: age and last_deficiency_no as a toy example.
Example 11.1: The details of the 50 selected ships are shown in Table 11.2. The
distribution of the features of the ships is shown in Figure 11.2.
In particular, we choose ‍K = 2‍, ‍K = 3‍, and ‍ K = 4‍ in the K-­means algorithm.
K-­means is implemented by scikit-­learn API [1]. It can easily be called by using the
following lines of codes:

‍ ‍

The corresponding clusters are shown in Figures 11.3–11.5.

Table 11.2   Ship information in the toy example

Ship ID age last_deficiency_no Ship ID age last_deficiency_no

1 8 1 2 5 0
3 22 5 4 8 0
5 13 2 6 14 0
7 20 13 8 11 5
9 10 0 10 14 1
11 9 5 12 16 6
13 19 4 14 19 9
15 6 0 16 8 4
17 6 9 18 11 17
19 14 1 20 16 4
21 19 5 22 38 8
23 5 1 24 9 4
25 12 0 26 13 1
27 18 8 28 10 1
29 11 0 30 9 11
31 8 0 32 23 17
33 9 0 34 4 1
35 7 1 36 20 3
37 2 4 38 9 3
39 15 2 40 14 8
41 4 1 42 19 3
43 13 2 44 4 0
45 7 1 46 16 0
47 6 0 48 17 0
49 8 4 50 18 4
 Cluster analysis  157

Figure 11.2   Distribution of the features of ships in the toy example

Figure 11.3   K-­means with ‍ K = 2‍for ship clustering

158  Machine learning and data analytics for maritime studies

Figure 11.4   K-­means with ‍K = 3‍for ship clustering

Figure 11.5   K-­means with ‍ K = 4‍for ship clustering

 Cluster analysis  159

In Figures  11.3–11.5, ships in different clusters are represented by different

symbols with different colors, and the centroids of the clusters are all in color orange
and the same symbol as the examples in the corresponding cluster of a larger size.
When ‍ K = 2‍, the centroid of cluster 1 is [18.9, 6.4], which means that the average
age of the ships in this cluster is 18.9 and the average last deficiency number of these
ships is 6.4. The centroid of cluster 2 is [8.6, 2.0]. It can be seen that the two clusters
of ships are quite different in both ship age and the last inspection state, and it can
also be anticipated that ships in cluster 1 have worse performance in the next PSC
inspection, as both older ships and worse performance in the last PSC inspection are
indicators of a higher ship risk.
When ‍K = 3‍, the centroids of clusters 1–3 are [16.3, 3.1], [7.5, 2.2], and [22.2,
12.8], respectively, which also show a positively correlated relationship between
ship age and the last deficiency number: as ship age increases from 7.5, to 16.3, and
to 22.2, the last deficiency number increases from 2.2, to 3.1, and to 12.8. Moreover,
as the last deficiency number of most ships is no more than 10 and ship age of most
ships is less than 20, the centroids of clusters 1 and 2 are closer to each other com-
pared to their distances from the centroid of cluster 3. Similar to the situation when
‍K = 2‍, it can also be expected that ships in cluster 3 have the highest risk, followed
by ships in cluster 1 and then by cluster 2.
When ‍ K = 4‍, the centroids of clusters 1–4 are [16.3, 11.9], [7.3, 1.9], [16.1, 2.4],
and [38.0, 8.0], respectively. Cluster 4 only has one ship, which is very old and is
far from the other ships. Then, it can be regarded as an outlier. For the other three
clusters, the general relationship between ship age and the last deficiency number
is similar to the situations when ‍ K = 2‍ and ‍K = 3‍: these two features are positively
correlated with each other. Furthermore, it can also be expected that ships in cluster
1 have the highest risk, while ships in cluster 2 have the lowest risk. In addition, as
ship age is densely and nearly evenly distributed between 0 and 20, the difference of
ship age value in the centroids of clusters 2 and 3 is much larger than that of clusters
3 and 1. Meanwhile, as there are several ships with last deficiency number more than
7, the difference of the last deficiency number of a ship in the centroids of clusters 2
and 3 is much smaller than that of clusters 3 and 1.
Based on the above three figures, it can also be seen that ‍K = 3‍is the most suit-
able value in K-­means for ship clustering in this toy example, as the features of the
ships are distinctive in different clusters, while the distribution of the clusters are
relatively stable.

Reference

[1] Pedregosa F., Varoquaux G., Gramfort A., et al. ‘Scikit-­learn: machine learning
in python’. Journal of Machine Learning Research. 2011;12:2825–30.
This page intentionally left blank
 Chapter 12
Classic and emerging approaches to solving
practical problems in maritime transport

This chapter aims to discuss classic and emerging approaches adopted in existing aca-
demic literature to address practical problems in maritime transport. First, widely stud-
ied practical problems in the maritime industry are summarized according to review
papers. Then, classic methods that are widely adopted by the related studies are intro-
duced. After that, data-­driven methods, as a typical type of emerging approaches, are
discussed. Especially, several examples of applying data-­driven models to address
prediction tasks in maritime studies are given. Finally, several issues regarding the
application of data-­driven models to address practical problems in maritime transport
are presented with various examples in port state control (PSC) provided.

12.1 Topics in maritime transport research

Maritime transport is the transportation of cargo over the global water transporta-
tion network, where seaports serve as the nodes and waterways are the links of the
network. Therefore, research topics in maritime transport academic research can be
divided into “shipping part” and “port part” according to Talley [1]. There are sev-
eral review papers on discussing the research themes, topics, and trends of the aca-
demic literature on maritime transportation, which we summarize as follows from
the aspects of the number of papers involved, the research areas and topics covered,
and main conclusions.

1. A series of review studies have been conducted by Woo et al. in 2011, 2012,
and 2013 on how seaports have been studied from 1980s to 2000s from
three perspectives: research themes and trends [2], research methodology
[3], and research collaboration [4]. The authors found that in 1980s, about
half of the papers in academic journals are published by Maritime Policy &
Management. In 1980s, more than one-­third of the related papers are still pub-
lished by Maritime Policy & Management, while more journals, and some of
them are newly established, started to publish papers on maritime transport,
such as Maritime Economics & Logistics, Journal of Transport Geography,
Transportation Research Part E: Logistics and Transportation Review, Marine
Policy, and Transport Policy. The authors divide port research into two main
162  Machine learning and data analytics for maritime studies

categories: government based and port authority/companies based, while the

former can be further divided into port policy and governance and reform, and
the latter can be further divided into management and strategy, competition
and performance, and port in supply chains. Regarding the research strate-
gies in port research, analytical and empirical strategies are the main research
strategy, where analytical strategy includes conceptual, mathematical, and sta-
tistical methods, and empirical strategy includes experimental and case studies
methods. In addition, the discipline bases in port research include econom-
ics, operations research, geography, strategic management, logistics, and mar-
keting. Regarding the trend of research collaboration, the authors conclude
that the research collaboration has substantially increased over the past three
decades and in particular publication via collaboration has become dominant
over the last decade.
2. Talley [1]: this study focuses on papers published in two leading maritime
transport journals: Maritime Policy & Management and Maritime Economics &
Logistics. Especially, papers published in Maritime Policy & Management from
2001 to 2012 are covered, and the most popular research topics of these papers
are shipping performance/management, shipping finance, port governance/
privatization, seafarers, and air emissions from shipping activities. Papers
published in Maritime Economics & Logistics during 2001–2012 are covered,
where the most popular topics are port performance, shipping performance/
management, shipping finance, port choice, and supply chain.
3. Shi and Li [5]: this study reviews 1292 papers on maritime transport published
in 19 transportation journals between 2000 and 2014. Among all the papers
covered, more than one-­third of them are published in Maritime Policy &
Management, followed by Maritime Economics & Logistics that publishes
about one-­fifth of the papers, and then by Transportation Research Part E,
International Journal of Shipping and Transport Logistics, and Journal of
Transport Geography. Regarding the research areas covered, nearly 51% of the
reviewed papers are in the area of shipping, about 43% are regarding port, and
about 6% are related to maritime fleet. The authors then summarize the research
topics under each of these three areas. Especially, the top five topics in shipping
research include shipping market, industry, freight rate, and economic impact;
operational management, mode choice, empty container; sailing speed, green
shipping, and environment-­related issues; shipping risk and maritime security;
and corporate management, stock returns, and performance. The top four topics
in port research include port management, service, performance, efficiency, and
competitiveness; terminal studies and berth allocation; port planning, develop-
ment, cluster, network, and economic impact, port governance, port policy, reg-
ulation, and legal issues; and ports in transport and supply chain. For maritime
fleet, most of the related papers are regarding shipbuilding, demolition, new
orders, and second-­hand ships and fleet structure, deployment, ownership, and
operation.
 Classic and emerging approaches to solving practical problems  163

12.2  R
 esearch methods and their specific applications to
maritime transport research

According to Shi and Li [5] and Yan et al. [6], seven methods are the most popular
in academic research on maritime transportation, whose explanations are shown in
Table 12.1.
From Table 12.1, it can be seen that the majority of the classic research meth-
ods are in a qualitative manner, including SIQO, case study, CCCQ, and literature
review. Those in a quantitative manner mainly aim to construct models using math-
ematical modeling, economic and econometric theories, statistical modeling, and
simulation methods to model or mimic the practical behavior of the system, with the
aim to gain operational and managerial insights to reduce costs or increase benefits
gained. The common point of the above models is that only when very specific
instructions are given to them can they function well. After a long time of develop-
ment and implementation, these classic methods as well as the results obtained from
them are widely recognized and used by both academic researchers and industrial
practitioners. However, it should also be mentioned that these methods are highly
dependent on long-­term practical experience and expert knowledge, while some
of them might be biased, and they might not be adapted and updated to the ever-­
changing environment.
A typical example of using classic methods to address a practical problem in PSC
is a ship selection scheme called ship risk profile (SRP) for high-­risk ships selection
introduced in Chapter 2, which is widely used by port states around the world for
higher risk ship selection. It is heavily based on expert knowledge, which can be seen
from the rough processing of the risk factors considered (e.g., only two states are con-
sidered for feature ship age: ships with age larger than 12, and ships with age no more
than 12), the fixed and subjective weighting points assigned to the states of the features
(e.g., ships with age larger than 12 are assigned with 2 weighting points, and 0, other-
wise), and the simple weighted-­sum manner for ship overall risk calculation. In addi-
tion, although ships can be quite different from each other, they are only divided into
three risk profiles by the SRP, namely, HRS (high-­risk ship), SRS (standard-­risk ship),
and LRS (low-­risk ship) from the highest risk to the lowest risk, and a fixed (while dif-
ferent) time window is attached to each risk profile. The time window attached to HRS
is longer than SRS, and is much longer than LRS. Then, a ship’s inspection priority
is highly dependent on the relationship between its last inspection time in the related
MoU (Memorandum of Understanding) and the time window attached to the ship: if
the period between the last inspection time and the current inspection is longer than
the upper bound of the time window attached to the ship considering its SRP, the ship
has high inspection priority; if the period is between the lower and upper bounds of the
time window attached, the ship has medium inspection priority; if the period is shorter
than the lower bound of the time window attached, the ship has low inspection priority.
Besides, even if ship flag and RO (recognized organization) performances are updated
annually and ship company performance is updated on a daily basis, feature process-
ing methods, weighting points attached to the feature states, inspection time windows
164  Machine learning and data analytics for maritime studies

Table 12.1  Summary of classic methods widely used in maritime transport

research [6]

Research methods Explanations

SIQO “S” represents survey, which is to collect information from a
selected sample of individuals by distributing and collecting
their responses to certain questions. The whole procedure
includes data collecting, aggregating, and analyzing.
“I” represents interview, which is a conversation aiming to gather
information and involves an interviewer, who is responsible
to coordinate the whole process and ask questions, and an
interviewee, who responds to the questions.
“Q” represents questionnaire, which is a research instrument
that consists of a set of questions with the aim to collect
information from respondents. It should be noted that
questionnaire and survey are related but different terms. Survey
contains a relatively wider range of action from question
design, data collection and analysis, and conclusion drawing.
This shows that a survey usually involves questionnaires, while
a single questionnaire is only one small part of survey.
“O” represents observation, which is a study of nonexperimental
situations where behavior is observed and recorded, so as to
generate either qualitative or quantitative results.
Economic modeling Using economic theories as well as quantitative and/or qualitative
models and techniques to analytically evaluate the cause and
effects of economic phenomenon.
MES “M” represents mathematical analysis, which aims to model a
nonmathematical situation, phenomenon, and the relationship
between the situations mathematically, e.g., by developing
mathematical models and proposing proper solution
approaches.
“E” represents econometric analysis, which aims to construct
a set of equations to provide quantitative explanations and
design tests to draw meaningful inferences from the samples to
the whole population.
“S” represents statistical analysis, which aims to collect and
analyze data based on statistical assumptions and tests so as
to draw meaningful inferences from the samples to the whole
population.
Case study An in-­depth study of the research object from comprehensive
perspectives to explore patterns and causes of its behaviors and
states.
CCCQ The first “C” represents conceptual analysis, which aims to break
down and analyze concepts into their constituent parts to gain a
better understanding of a particular issue.
The second “C” represents content analysis, which aims to
quantify and analyze the presence of certain words, themes, or
concepts within some given quantitative data.
(Continues)
 Classic and emerging approaches to solving practical problems  165

Table 12.1  Continued

Research methods Explanations
The third “C” represents comparative analysis: which aims to
compare two or more processes, documents, data sets, and
other objectives.
“Q” represents qualitative analysis, which uses subjective
judgments to analyze the research objective based on
nonquantifiable information.
Literature review An overview and evaluation of the available literature in a given
subject of area.
Simulation An imitation of the operation of a real-­world process or system
over time.

attached to the risk profiles, and ship inspection priority are not updated considering
the changing conditions and factors, such as the evolution of ship conditions and the
influences of the COVID-­19 on ship behaviors and the PSC inspection mode.
To reduce the adverse influences of expert knowledge and subjective judgment
on making assumptions and constructing models and to take the ever-­changing
environment into account, data-­driven models as an emerging approach are rapidly
developed and used to address practical problems in maritime transport research. As
the name suggests, data-­driven models are highly dependent on the data collected,
so as to dynamically explore useful information from these data. ML (machine
learning) models, which allow computer to learn from data facilitated by indirect
programming, are a typical type of data-­driven models. Especially, an overview of
data-­driven models is given in Chapter 3 and the key elements of data-­driven models
are summarized in Chapter 1 of this book. There are some initial trials of applying
data-­driven models to deal with practical issues in maritime transportation, which
we briefly list as follows [6]:

1. Ship trajectory prediction (in the category of shipping research): it aims to pre-
dict the trajectory of a moving vessel in the near future by using historical time
series trajectory points of the vessel, which is the foundation of vessel naviga-
tion risk analysis. Ship sailing trajectory is constructed from vessel automatic
identification systems (AIS), where one sailing record comprising of vessel
dynamic sailing information is proposed every tens of seconds to a few min-
utes. Popular data-­driven models in existing literature for ship trajectory predic-
tion include SVM (support vector machine) [7], extreme learning machine [8],
k-­nearest neighbors [9], autoencoder [10], and long short-­term memory net-
works [11, 12].
2. Ship risk prediction and safety management (in the category of shipping
research): two research topics are widely studied under this research stream:
prediction of the occurrence probability of vessel accidents, and the analysis of
regional and global ship accident statistics and accident reports. In the former
research topic, AIS is the main data resource used to generate ship historical
166  Machine learning and data analytics for maritime studies

trajectory and predict ship trajectory in the near future for risk assessment and
analysis of ship collision [13–15], ship grounding accidents [16], fire accidents
[17], and seafarers’ nonfatal injuries [18]. In the latter research topic, mari-
time accident statistics and reports published by the IMO (Internation Maritime
Organization) and local marine departments/institutions are the main data
resource. BNs are the most popular model to analyze the relations among the
risk factors and between the risk factors and the prediction target. The related
studies can be found in References [19–22]. In addition, K-­means is also applied
to analyze maritime accidents [23, 24].
3. Ship inspection planning (in the category of shipping research): the related
research topics are lying in four categories: ship risk prediction, onboard inspec-
tion sequence optimization, influencing factors on the inspection results of PSC,
and PSC inspections’ influence. Research data are mainly from the inspection
data published by individual MoUs and ship specification data provided by
online commercial databases. Widely used data-­driven models in PSC-­related
research include BNs [25, 26], tree-­based models [27–29], SVMs [30, 31], and
association rule learning [32–34].
4. Ship energy efficiency prediction (in the category of shipping research): the
related research aims to predict ship fuel consumption rate (i.e., hourly or daily
fuel consumption amount) under various conditions using features of ship sail-
ing information and the surrounding environments. Such ship- and environment-­
related features are collected from manually filled ship sailing records like noon
reports or automatically collected from onboard sensors like fuel consumption
sensors, global positioning system receivers, shaft power testers, wind speed
sensors, and water depth sonars. ANNs are the most widely used fuel con-
sumption rate prediction model, which can be found in References [35–38].
Additionally, tree-­based models [39, 40], LASSO (least absolute shrinkage
and selection operator) regression [41, 42], ridge regression [43, 44], and SVR
(support vector regression) [44] are also popular for ship energy efficiency pre-
diction. Then, the predicted ship energy efficiency indicators, with fuel con-
sumption rate as the representative, are used to guide voyage management and
optimization by means of speed optimization, trim optimization, weather rout-
ing, and their combinations.
5. Ocean freight market condition prediction (in the category of shipping
research): two indicators of shipping market condition are widely used as the
target of ocean freight market condition prediction, namely, baltic dry index
(BDI), which is regarded as the leading indicator of economic activity and the
barometer of the dry bulk shipping market, and sea freight rate, which is the
requested price to transport cargo from the origin to the destination by sea,
mainly determined by the weight and volume of the cargo and the distance
between the origin and destination. For the prediction of BDI, ANNs (artifical
neural networks) and SVRs are the most popular methods, and the adoption of
ANNs can be seen in References [45–47], while the adoption of SVRs can be
seen in References [48, 49]. For the prediction of sea freight rate, ANNs are
also a popular method, which can be seen in References [50–52]. Other ML
 Classic and emerging approaches to solving practical problems  167

models are also used and compared for sea freight rate prediction, as can be
seen in [53, 54].
6. Ship destination port and arrival time at port prediction (in the category of port
research): the related studies aim to predict the destination of a voyage and its
arrival time. As such prediction is highly dependent on ship historical voyage
information and the information of the current voyage, AIS is a widely used
data source. Ship arrival time to the destination is also influenced by the sea
and weather conditions along the way, as well as the port operating policies and
conditions (e.g., the just-­in-­time arrival policy for ships to save fuel consump-
tion and reduce greenhouse gas emissions). Therefore, marine weather forecast
and port statistics are also important data resources. It should be mentioned
that although the destination of the current voyage is required to be reported
via the AIS by vessel captains, it is widely believed that such report is highly
unreliable with complex causes as discussed by Yang et al. [55]. For ship des-
tination prediction, the methods can be either turning point based or trajectory
extraction based. As both tasks are nontrivial, especially when there are require-
ments on prediction accuracy, various ML-­based models are developed, such as
a framework based on anomaly detection and route prediction [56], DBSCAN
(density-­based spatial clustering of applications with noise) for turning region
clustering and identification combined with ANN for next turning point predic-
tion [57]. For literature in the second category, the prerequisite of arrival time
prediction is to know the destination. Therefore, there are several studies that
jointly predict vessel destination and the vessel arrival time, which can be found
in References [58–60].
7. Port condition prediction (in the category of port research): this stream of
research aims to predict port cargo volume and port traffic volume, where the
related data are from port statistics and bill of lading. Port cargo volume is
predicted by ANNs combined with time series methods [24] and by SVRs [61].
Port traffic volume at the Shanghai port is predicted by ANNs in Reference [62].

12.3 Issues of adopting data-driven models to address problems

in maritime transportation

As introduced in Chapter 4, the procedure of adopting data-­driven models to address

a practical problem in maritime transportation starts from understanding the appli-
cation scenarios and specification of the current problem as well as the feasibility
of solving it. Then, related data are collected, which will be preprocessed and pro-
cessed by various feature engineering techniques, including data cleaning, feature
extraction, feature preprocessing, feature encoding and scaling, and feature selec-
tion, whose aim is to allow developing a proper and efficient data-­driven model
that can solve the problem more effectively. After that, a model/several models are
constructed, evaluated, and refined while considering the properties and require-
ments of the practical problem to be solved, and then the best model on hand is
168  Machine learning and data analytics for maritime studies

finally selected and handed over to practical users to deal with the current problem.
The users can be technicians who are familiar with the underlying of the model or
practitioners who are unfamiliar with the theories of the model. During the practical
usage of the model, ideally, its input, output, and settings should be constantly evalu-
ated and updated, so as to better handle the ever-­changing environment.
In the above process, four layers are involved and interwoven with each other,
namely, data, model, user, and target. To develop proper and efficient data-­driven
models, there are issues within these four layers that should be paid close attention
to which have been discussed by Yan et al. [6]. We discuss these issues here in detail
with several new points added in the following subsections.

12.3.1 Data
Regarding the data used for model construction, one critical issue is that the data
used, including the input features and output target(s), should be highly compat-
ible with the focus of the current problem. To be specific, the features used should
possess the properties of “relevance” and “comprehensiveness”: “relevance” means
that they are indeed related to/can contribute to the prediction of the target, and
“comprehensiveness” means that the features used cover most of the indicators that
affect the target.
Issues regarding data relevance can be assessed from qualitative perspective
and quantitative perspective. Qualitatively, features selected should be of interest to
model users, which is demonstrated in the following Example 12.1 of PSC.
Example 12.1: To predict the risk of a ship in a PSC inspection, it is common
that ship characteristics and indicators of its historical inspection performance, such
as ship dimensions (i.e., length, depth, beam, and draft), ship’s flag, RO, and com-
pany performance, the date and performance of a ship’s last PSC inspection, etc.,
are used as the input to develop prediction models, as can be seen in References
[27, 28, 30, 31]. These features are relevant to ship’s structural and operational
conditions, navigation safety, and safe management, which are also key points in
PSC. Therefore, it is justifiable to consider these factors when developing ship risk
prediction models. In contrast, sailing information of the current voyage, such as
the sailing route, average speed, freight rate, and charter mode, should deserve less
attention since obtaining such information is costly or even impossible, while the
port authorities cannot benefit much from the above information when evaluating
ship risk levels.
Quantitatively, feature selection can be dealt with by relevance calculation and
evaluation model development. The relevance calculation approach aims to evaluate
the interrelationship between the features and the target. For example, in the step of
feature selection in feature engineering, several approaches of feature selection can
be used to identify the most relevant features from all candidate features by calculat-
ing the relation between the features and the target(s) as introduced in section 1.2.4
of Chapter 4. Such approaches include Pearson’s correlation coefficient, mutual
information, and recursive feature elimination. In addition, features can be selected
based on regularization and feature importance, and both can be integrated in an ML
 Classic and emerging approaches to solving practical problems  169

algorithm. In particular, the former is by shrinking the coefficients, or weights, of the

features that have very little impact on the prediction target to zero, so as to exclude
this feature from the predicting process. The latter can be mainly achieved by tree-­
based models and permutation, as feature importance can be automatically gener-
ated by a tree-­based model evaluated by the impurity reduction in node splitting by
the features, while features that lead to higher prediction error in the permutation
process should be more important in the permutation method.
In contrast, comprehensiveness of the features used can be much harder to eval-
uate, and it is highly dependent on and restricted by the available data sets that can
be obtained. To consider more comprehensive features, data fusion is widely used
to combine different data resources. Example 12.2 is provided as an example of data
fusion in PSC inspection.
Example 12.2: In the Tokyo MoU where the port of Hong Kong belongs to,
an online database called Asia Pacific Computerized Information System (APCIS)
provides the detailed inspection records, where the record of an individual inspec-
tion includes ship basic information (i.e., ship name, ship type, IMO number,
MMSI (maritime mobile service identity), call sign, classification society, flag, date
keel laid, dead weight, and gross tonnage, ship’s flag/RO/company performance)
as well as the inspection details. The basic ship information can no doubt provide
insights on the prediction of ship risk. Nevertheless, they cannot cover a full picture
of ship conditions where the PSC inspectors might be interested in. For example,
as shown by the econometric results of Cariou and Wolff [63], ships with worse
historical PSC inspection performance are more likely to change their flag states.
Therefore, it can be expected that taking ship flag change times between the period
of keel laid date and the current inspection can increase prediction accuracy. Such
information is provided by the World Register of Ship database, which can be com-
bined with the APCIS database by ship IMO number. In addition, as ship inspec-
tion priority is highly related to its last inspection time as well as its performance in
previous PSC inspections, taking factors regarding previous inspections, including
but not limited to the last inspection time and the deficiency and detention condi-
tions, is highly likely to improve the performance of the prediction model. Such
information can be searched from APCIS by using a ship’s IMO number, and can
be combined with the original inspection records collected from APCIS by ships’
IMO numbers.

12.3.2 Model
When developing data-­driven models to solve a practical problem in maritime trans-
portation, it is of utmost importance to understand and clarify the properties of the
problem itself as well as the proper metrics to evaluate its performance. One basic
step is to identify the nature of the current problem, i.e., to decide the current prob-
lem is a classification, regression, clustering, or association rule mining task. Then,
models falling within certain category should be tried and evaluated. In this process,
how to select an effective model as well as its evaluation metrics should be carefully
planned. We use the following Example 12.3 to show the related issues.
170  Machine learning and data analytics for maritime studies

Example 12.3: Suppose that we need to address the problem of ship detention
prediction in PSC. The nature of this problem is classification, where a detained ship
is predicted to be “1” and a non-­detained ship is predicted to be “0.” The main char-
acteristic of this problem is that the data set is highly imbalanced regarding the pre-
diction target: the number of non-­detained ships is much larger than that of detained
ships. For example, the annual detention rate at the Hong Kong port ranges from
2.96 to 3.67% between 2015 and 2019 as shown in the 2020 annual report of Tokyo
MoU [64]. This means that among all 100 foreign visiting ships to the Hong Kong
port that are inspected, no more than 3.67 of them are actually detained on average,
while up to 97.04 of them are actually not detained on average. If standard classi-
fication models are directly applied, no matter how well they perform on balanced
data sets, it can be anticipated that their performance for ship detention prediction
would be relatively bad, or the prediction results cannot make sense (e.g., predicting
all the inspected ships to be non-­detained leading to a very high accuracy score, but
the prediction is meaningless).
To take the imbalanced property of the data set into account, several strategies
can be adopted. For example, oversampling (i.e., randomly selecting a certain num-
ber of detained ships from the original data set and then input the selected ships into
the original data set to form a balanced one) or subsampling (i.e., randomly select-
ing a subset of non-­detained ships from the original data set and then form a new
data set containing all detained ships and the subset of the non-­detained ships) can
be applied before model construction. Moreover, specialized ML models suitable to
address imbalanced classification tasks can also be used, such as balanced random
forest and anomaly detection models.
Another major issue regarding data-­driven models is that most of the ML
models are of black-­box nature, meaning that the model work mechanism and the
prediction results cannot be understood by the developers and users. There are of
course some exceptions, such as LR (linear regression), DT (decision tree), and
BN (Bayesian network) models. Nevertheless, as these models are not sophisti-
cated enough, they may have bad performance when addressing certain complex
problems, and thus other sophisticated ML models of black-­box nature have to
be adopted. The black-­box property of data-­driven models would reduce their
acceptance and credibility of the practitioners, and thus restrict their applicabil-
ity. Furthermore, it is well known that when models are not conceived with self-­
explanatory characteristics, they may engender pitfalls. In recent years, a concept
of “explainable artificial intelligence,” or XAI for short, is proposed and rapidly
developed, which aims to open up the black boxes of ML models and shed light
on the model working mechanisms as well as the justification of the results gener-
ated. Details of the concepts, techniques, and applications of XAI will be covered
in later sections of this part.

12.3.3 User
It is widely known that the maritime industry is relatively conservative and tradi-
tional, in which the decision support tools used by maritime practitioners are mainly
 Classic and emerging approaches to solving practical problems  171

based on expert knowledge or classic theories for a long time, and the stakeholders
are reluctant to share the information and data with each other. Consequently, with-
out the support of sufficient data, especially high-­quality data, the digitization pro-
cess of the maritime industry is relatively slow, where contemporary approaches,
especially data-­driven models, are not that popular compared to other industries
such as financial technology and healthcare industry. Therefore, data-­driven appli-
cations are still in their infancy, and it is also noted that applying data-­driven mod-
els to maritime transport industry requires replacing the current (naive) decision
support tools, which the decision makers are quite familiar with more sophisticated
models that the decision makers may be unfamiliar with. Systems based on such
models are more like black boxes to the users, with brand new graphical user inter-
face (GUI) and obscure working mechanism. In this process, one guiding principle
is to avoid imposing any unnecessary burden or pressure on the users during their
usage.
However, this is not a trivial task as the developers and users of a data-­driven-­
based model or system in the maritime industry usually have quite different back-
grounds. On the one hand, model and system developers are usually engineering or
researchers specializing in data analytics and with weak shipping background. The
data-­driven models developed usually have multifarious inputs and outputs as well
as complex structures, which means that the data processing and model operation
can be sophisticated. On the other hand, model and system users are usually prac-
titioners of the maritime industry, like government officials, managers, and techni-
cians in shipping companies and classification societies, and crew members, who
may lack expertise in data analytics and system operation. It is therefore crucial
to avoid requiring model users to input extra data or too much information to the
models as well as conducting too many extra operations when using the systems.
Ideally, a system with straightforward, concise, and friendly GUI as well as easy
human–computer interactions should be developed. The following is an example
of a high-­risk ship selection decision support system in PSC developed for the
Marine Department of Hong Kong by researchers from the Hong Kong Polytechnic
University.
Example 12.4: This system is named by AI for PSC at Hong Kong*, which
automatically downloads the information of in port ships and due to arrival ships
at the Hong Kong port about every 20 minutes, and then applies AI (artificial
intelligence)-­based models to predict ship risk, which is represented by the com-
bination of the number of deficiencies and the detention probability of each ship
to facilitate the Marine Department to select ships with the highest risks for
inspection. Especially, the displayed items as well as their formats on the website,
including not only the outputs of the AI models but also the ships’ identify infor-
mation, type, flag, last port of call, agent, current location, etc., are compliant with
the preference of the decision makers. This means that the system users do not

* 
https://sites.google.com/site/wangshuaian/research/
172  Machine learning and data analytics for maritime studies

need to do anything to use the system, and they can acquire the information they
are interested in for high-­risk ship selection directly from the system. Therefore,
the newly developed ship selection system based on data-­driven models is in com-
pliance with the users’ past working habits and preference, and thus does not pose
extra operational burdens to them.
Another issue in this category is regarding data-­driven models’ applicabil-
ity: as such models are developed using data collected from certain region and
entities, while different regions and entities might be quite different from each
other, they do not possess the property of universal applicability. For example,
the ship risk prediction models developed in this book all use PSC inspection
records at the Hong Kong port, and thus they may be hard to be applied to ports
outside of the Tokyo MoU, as different ship selection methods are used, and
even to the other ports within the Tokyo MoU, as the expertise and background
of the inspectors in different ports might vary a lot. Therefore, if one wishes to
apply the ship risk prediction models developed in this book to ports other than
the Hong Kong port, new prediction models using the corresponding data sets
should be developed, and more advanced methods, such as transfer learning, can
also be used.

12.3.4 Target
To promote the acceptance and application of data-­driven models, the predicted out-
put, i.e., the target values given by the data-­driven models, should comply with ship-
ping domain knowledge. Otherwise, it is highly likely that the users would doubt the
accuracy and fairness of the prediction models developed. The following Example
12.5 shows this issue in high-­risk ship selection in PSC.
Example 12.5: In many existing studies on ship risk prediction, ship flag,
RO, and company performances are considered as input features to the ship risk
prediction models, which can be found in References [27–29]. Roughly speaking,
these three factors are calculated based on the performance of the ships under
their management in the previous three years, and can thus influence their reputa-
tion and the inspection priority of their ships in future PSC inspection. Therefore,
responsible ship flag, RO, and company put many efforts to maintain their ships
to be in satisfactory conditions, and they hope that their hard efforts can gain them
good reputation and thus a lower inspection priority of their vessels. Therefore,
it is justifiable to say that for two ships with all the other conditions being equal,
the ship with better flag, RO, or company performance than the other ship should
be predicted to be in a lower risk (i.e., less number of deficiencies or lower deten-
tion probability), and thus is less likely to be selected for inspection. If the pre-
dicted targets given by a data-­driven model conflict with such shipping domain
knowledge, the decision makers, as well as the ships’ flags, ROs, and companies
might hardly trust the predictions. Therefore, developers of the data-­driven mod-
els should take such shipping domain knowledge into account in model develop-
ment to guarantee model fairness. Details of this issue will be covered in the later
chapters of this part.
 Classic and emerging approaches to solving practical problems  173

To conclude, issues that deserve more attention in developing data-­driven mod-

els to address practical problems in maritime transport as discussed above are sum-
marized as follows:

1. data: relevance and comprehensiveness;

2. model: model selection and evaluation, black-­box nature of ML models;
3. users: putting no extra burden on users, model universal applicability;
4. targets: considering shipping domain knowledge.

It should be noted that there are still a long way to go to develop and apply
systems based on data-­driven models in the traditional and conservative mari-
time industry, and the issues covered are only a small part of the whole picture
that need more attention and efforts.

References

[1] Talley W.K. ‘Maritime transportation research: topics and methodologies’.

Maritime Policy & Management. 2013;40(7):709–25.
[2] Woo S.-H., Pettit S., Beresford A., Kwak D.-W. ‘Seaport research: a dec-
adal analysis of trends and themes since the 1980s’. Transport Reviews.
2012;32(3):351–77.
[3] Woo S.-H., Pettit S.J., Kwak D.-W., Beresford A.K.C. ‘Seaport research:
a structured literature review on methodological issues since the 1980s’.
Transportation Research Part A. 2011;45(7):667–85.
[4] Woo S.-H., Kang D.-J., Martin S. ‘Seaport research: an analysis of re-
search collaboration using social network analysis’. Transport Reviews.
2013;33(4):460–75.
[5] Shi W., Li K.X. ‘Themes and tools of maritime transport research during
2000-­2014’. Maritime Policy & Management. 2017;44(2):151–69.
[6] Yan R., Wang S., Zhen L., Laporte G. ‘Emerging approaches applied to mari-
time transport research: past and future’. Communications in Transportation
Research. 2021;1:100011.
[7] Liu X., He W., Xie J., Chu X. ‘Predicting the trajectories of vessels using
machine learning’. 5th International Conference on Control, Robotics and
Cybernetics (CRC); Wuhan, China, IEEE, 2020. pp. 66–70. Available from
https://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=9253444
[8] Mao S., Tu E., Zhang G., et al. ‘An automatic identification system (AIS)
database for maritime trajectory prediction and data mining’. Proceedings of
ELM-­2016; Springer, 2018. pp. 241–57.
[9] Virjonen P., Nevalainen P., Pahikkala T., Heikkonen J. ‘Ship movement predic-
tion using k-­NN method’. 2018 Baltic Geodetic Congress (BGC Geomatics);
Olsztyn, 2020. pp. 304–09.
[10] Murray B., Perera L.P. ‘A dual linear autoencoder approach for vessel trajectory
prediction using historical AIS data’. Ocean Engineering. 2020;209:107478.
174  Machine learning and data analytics for maritime studies

[11] Huang Y., Chen L., Chen P., Negenborn R.R., van Gelder P.H.A.J.M. ‘SHIP col-
lision avoidance methods: state-­of-­the-­art’. Safety Science. 2020;121:451–73.
[12] Li W., Zhang C., Ma J., Jia C. ‘Long-­term vessel motion predication by mod-
eling trajectory patterns with AIS data’. 2019 5th International Conference on
Transportation Information and Safety (ICTIS); Liverpool, IEEE, 2020. pp.
1389–94.
[13] Gang L., Ma J., Yao J. ‘Decision-­making of vessel collision avoidance based
on support vector regression’. ICAIIS 2021; Chongqing, New York, NY,
2021. pp. 1–6. Available from https://dl.acm.org/doi/proceedings/10.1145/​
3469213
[14] Gao M., Shi G.Y., Liu J. ‘Ship encounter azimuth MAP division based on
automatic identification system data and support vector classification’. Ocean
Engineering. 2020;213:107636.
[15] Abebe M., Noh Y., Seo C., Kim D., Lee I. ‘ Developing a SHIP collision
risk index estimation model based on dempster-­shafer theory’. Applied Ocean
Research. 2021;113:102735.
[16] Wu B., Yan X., Yip T.L., Wang Y. ‘A flexible decision-­support solution
for intervention measures of grounded ships in the Yangtze River’. Ocean
Engineering. 2017;141:237–48.
[17] Wu B., Tang Y., Yan X., Guedes Soares C. ‘Bayesian network modelling for
safety management of electric vehicles transported in ropax ships’. Reliability
Engineering & System Safety. 2017;209:107466.
[18] Zhang G., Thai V.V., Law A.W.K., Yuen K.F., Loh H.S., Zhou Q.
‘Quantitative risk assessment of seafarers’ nonfatal injuries due to occu-
pational accidents based on Bayesian network modeling’. Risk Analysis.
2020;40(1):8–23.
[19] Wang L., Yang Z. ‘Bayesian network modelling and analysis of accident
severity in waterborne transportation: a case study in China’. Reliability
Engineering & System Safety. 2018;180:277–89.
[20] Fan S., Blanco-­Davis E., Yang Z., Zhang J., Yan X. ‘Incorporation of hu-
man factors into maritime accident analysis using a data-­driven Bayesian net-
work’. Reliability Engineering & System Safety. 2020;203:107070.
[21] Fan S., Yang Z., Blanco-­Davis E., Zhang J., Yan X. ‘Analysis of maritime
transport accidents using Bayesian networks’. Proceedings of the Institution
of Mechanical Engineers, Part O. 2020;234(3):439–54.
[22] Li B., Lu J., Lu H., Li J. ‘Predicting maritime accident consequence scenarios
for emergency response decisions using optimization-­based decision tree ap-
proach’. Maritime Policy & Management. 2020:1–23.
[23] Lema E., Papaioannou D., Vlachos G.P. ‘Investigation of coinciding shipping
accident factors with the use of partitional clustering methods’. PETRA ’14;
Rhodes Greece, New York, NY, 2014. pp. 1–4. Available from https://dl.acm.​
org/doi/proceedings/10.1145/2674396
[24] Zhang Y., Sun X., Chen J., Cheng C. ‘Spatial patterns and characteristics
of global maritime accidents’. Reliability Engineering & System Safety.
2021;206:107310.
 Classic and emerging approaches to solving practical problems  175

[25] Yang Z., Yang Z., Yin J. ‘Realising advanced risk-­based Port state control
inspection using data-­driven Bayesian networks’. Transportation Research
Part A. 2018;110:38–56.
[26] Wang S., Yan R., Qu X. ‘Development of a non-­parametric classifier: effec-
tive identification, algorithm, and applications in Port state control for mari-
time transportation’. Transportation Research Part B. 2019;128:129–57.
[27] Yan R., Wang S., Fagerholt K. ‘A semi- “ smart predict then optimize ” (semi-­
SPO) method for efficient SHIP inspection’. Transportation Research Part B.
2020;142:100–25.
[28] Yan R., Wang S., Cao J., Sun D. ‘Shipping domain knowledge informed pre-
diction and optimization in Port state control’. Transportation Research Part
B. 2021;149:52–78.
[29] Yan R., Wang S., Peng C. ‘An artificial intelligence model considering data
imbalance for SHIP selection in Port state control based on detention prob-
abilities’. Journal of Computational Science. 2021;48:101257.
[30] Xu R.-F., Lu Q., Li W.-J., Li K.X., Zheng H.-S. ‘Presented at 2007
International Conference on Machine Learning and Cybernetics’. Hong
Kong.
[31] Wu S., Chen X., Shi C., Fu J., Yan Y., Wang S. ‘Ship detention prediction
via feature selection scheme and support vector machine (SVM)’. Maritime
Policy & Management. 2022;49(1):140–53.
[32] Tsou M.C. ‘Big data analysis of port state control ship detention database’.
Journal of Marine Engineering & Technology. 2019;18(3):113–21.
[33] Chung W.-H., Kao S.-L., Chang C.-M., Yuan C.-C. ‘Association rule learning
to improve deficiency inspection in Port state control’. Maritime Policy &
Management. 2020;47(3):332–51.
[34] Yan R., Zhuge D., Wang S. ‘Development of two highly-­efficient and in-
novative inspection schemes for PSC inspection’. Asia-­Pacific Journal of
Operational Research. 2021;38(3):2040013.
[35] Pedersen B.P., Larsen J. ‘Prediction of full-­scale propulsion power using ar-
tificial neural networks’. Proceedings of the 8th International Conference
on Computer and IT Applications in the Maritime Industries (COMPIT’09);
Budapest: Hungary, 2009. pp. 10–12.
[36] Petersen J.P. Mining of SHIP operation data for energy conservation. DTU
Informatics; 2012.
[37] Rudzki K., Tarelko W. ‘A decision-­making system supporting selection of
commanded outputs for A ship’s propulsion system with A controllable pitch
propeller’. Ocean Engineering. 2016;126:254–64.
[38] Du Y., Meng Q., Wang S., Kuang H. ‘Two-­phase optimal solutions for
ship speed and trim optimization over a voyage using voyage report data’.
Transportation Research Part B. 2019;122:88–114.
[39] Peng Y., Liu H., Li X., Huang J., Wang W. ‘Machine learning method for
energy consumption prediction of ships in Port considering green ports’.
Journal of Cleaner Production. 2020;264:121564.
176  Machine learning and data analytics for maritime studies

[40] Yan R., Wang S., Du Y. ‘Development of a two-­stage shipfuel consumption

prediction and reduction model for a dry bulk ship’. Transportation Research
Part E. 2020;138:101930.
[41] Wang S., Ji B., Zhao J., Liu W., Xu T. ‘Predicting ship fuel consumption based
on LASSO regression’. Transportation Research Part D. 2018;65:817–24.
[42] Gkerekos C., Lazakis I., Theotokatos G. ‘Machine learning models for pre-
dicting ship main engine fuel oil consumption: a comparative study’. Ocean
Engineering. 2019;188:106282.
[43] Soner O., Akyuz E., Celik M. ‘Statistical modelling of ship operational per-
formance monitoring problem’. Journal of Marine Science and Technology.
2019;24(2):543–52.
[44] Uyanık T., Karatuğ Ç., Arslanoğlu Y. ‘Machine learning approach to SHIP
fuel consumption: a case of container vessel’. Transportation Research Part
D. 2020;84:102389.
[45] Leonov Y., Nikolov V. ‘A wavelet and neural network model for the pre-
diction of dry bulk shipping indices’. Maritime Economics & Logistics.
2012;14(3):319–33.
[46] Şahin B., Gürgen S., Ünver B. ‘Forecasting the baltic dry index by using an
artificial neural network approach’. Turkish Journal of Electrical Engineering
& Computer Sciences. 2021;26(3):1673–84.
[47] Bae S.H., Lee G., Park K.S. ‘A baltic dry index prediction using deep learning
models’. Journal of Korea Trade. 2021;25(4):17–36.
[48] Han Q., Yan B., Ning G., Yu B. ‘Forecasting dry bulk freight index with im-
proved SVM. Math Probl Eng. 2014;2014:1–12.
[49] Bao J., Pan L., Xie Y. ‘A new BDI forecasting model based on support vec-
tor machine’. Presented at 2016 IEEE Information Technology, Networking,
Electronic and Automation Control Conference (ITNEC); Chongqing, China.
IEEE,
[50] Cv S., HJv M., Breitner M.H. ‘Spot and freight rate futures in the tanker
shipping market: short-­term forecasting with linear and non-­linear methods’.
Presented at In: Operations Research Proceedings 2012; Springer,
[51] Santos A.A.P., Junkes L.N., Pires Jr F.C.M. ‘Forecasting period charter rates
of VLCC tankers through neural networks: a comparison of alternative ap-
proaches’. Maritime Economics & Logistics. 2014;16(1):72–91.
[52] Yang Z., Mehmed E.E. ‘Artificial neural networks in freight rate forecasting’.
Maritime Economics & Logistics. 2019;21(3):390–414.
[53] Ubaid A., Hussain F.K., Charles J. ‘Machine learning-­based regression mod-
els for price prediction in the Australian container shipping industry: case
study of asia-­oceania trade Lane’. In: International Conference on Advanced
Information Networking and Applications; Springer, 2020. pp. 52–59.
[54] Næss P.A. Investigation of multivariate freight rate prediction using machine
learning and AIS data. NTNU; 2018.
[55] Yang D., Wu L., Wang S. ‘Can we trust the AIS destination Port information
for bulk ships? –implications for shipping policy and practice’. Transportation
Research Part E. 2021;149:102308.
 Classic and emerging approaches to solving practical problems  177

[56] Pallotta G., Vespe M., Bryan K. ‘Vessel pattern knowledge discovery from
AIS data: a framework for anomaly detection and route prediction’. Entropy.
2018;15(12):2218–45.
[57] Daranda A. ‘Neural network approach to predict marine traffic’. Baltic
Journal of Modern Computing. 2016;4(3):483.
[58] Amariei C., Diac P., Onica E., Roşca V. ‘Cell grid architecture for maritime
route prediction on AIS data streams’. DEBS ’18; Hamilton, New Zealand,
New York, NY, 2018. pp. 202–04. Available from https://dl.acm.org/doi/
proceedings/10.1145/3210284
[59] Bodunov O., Schmidt F., Martin A., Brito A., Fetzer C. ‘R-­time destination
and eta prediction for maritime traffic’. DEBS ’18; Hamilton, New York, NY,
2018. pp. 198. Available from https://dl.acm.org/doi/proceedings/10.1145/​
3210284
[60] Nguyen D.D., Le Van C., Ali M.I. ‘Vessel destination and arrival time pre-
diction with sequence-­to-­sequence models over spatial grid’. DEBS ’18;
Hamilton, New York, NY, 2018. pp. 217–20. Available from https://dl.acm.​
org/doi/proceedings/10.1145/3210284
[61] Ruiz-­Aguilar J.J., Moscoso-­López J.A., Urda D., González-­Enrique J.,
Turias I. ‘A clustering-­based hybrid support vector regression model to
predict container volume at seaport sanitary facilities’. Applied Sciences.
2020;10(23):8326.
[62] Wang S., Wang S., Gao S., Yang W. ‘Daily SHIP traffic volume statistics and
prediction based on automatic identification system data’. 9th International
Conference on Intelligent Human-­ Machine Systems and Cybernetics
(IHMSC); Hangzhou, IEEE, 2017. pp. 149–54.
[63] Cariou P., Wolff F.C. ‘Do Port state control inspections influence flag-­and
class-­hopping phenomena in shipping?’. Journal of Transport Economics and
Policy (JTEP). 2011;45(2):155–77.
[64] Annual report. Tokyo: Tokyo MOU. Available from http://www.tokyo-mou.​
org/doc/ANN19-f.pdf
This page intentionally left blank
 Chapter 13
Incorporating shipping domain knowledge into
data-­driven  models

As mentioned in section 12.3.4 of hapter 12, one of the issues regarding the target of
applying data-­driven models to solve practical problems in maritime transportation
is that the predicted target may not comply with shipping domain knowledge. The
term “domain knowledge” refers to rules and common senses widely believed by
the practitioners. Domain knowledge is based on the practitioners’ understanding
of the disciplines and activities of the industry, and is gained from longtime experi-
ence of the practitioners in this industry as well as their own expertise, professions,
specializations, and judgment. For example, in the maritime industry, regarding the
activity of ship selection for inspection by PSC (port state control), it is generally
believed that given all other conditions being equal, an older ship would have a
larger number of deficiencies than a younger ship. Regarding ship fuel consumption
prediction, ship sailing speed is the most significant determinant, and it is widely
believed that a ship’s fuel consumption rate is proportional to its sailing speed to the
power of ‍˛ = 3‍, i.e., ‍r / ˇ  v˛,‍ where ‍r ‍is the hourly or daily fuel consumption, ‍ˇ‍
is a coefficient, and ‍v ‍is the average sailing speed. In practice, ‍ ˛‍can be higher than 3,
especially for large vessels like container ships where it can be 4, 5, or even higher.
In the former example, if there is a ship risk prediction model that gives opposite
prediction results, i.e., a younger ship has more deficiencies than an older ship under
the condition that the other features of these two ships are identical, the prediction
results as well as the prediction model are expected to be hardly accepted or used
by the port state officers, because they may conclude that the proposed model is
inaccurate and unfair based on such prediction. Similarly, in the latter example, if
the predicted ship fuel consumption rate does not take such convex and increasing
relationship with ship sailing speed, it may not convince the users. Given the condi-
tion that practitioners in the conservative while classic maritime industry might be
reluctant to replace the current rule-­based decision support systems with data-­driven
ones, it is thus of vital importance to make sure that the prediction results given by
the data-­driven models to address practical problems comply with the corresponding
shipping domain knowledge. Otherwise, the practitioners are more likely to be very
skeptical of the models together with their results, and thus are not willing to use
them to assist their decision-­making.
180  Machine learning and data analytics for maritime studies

One way to guarantee that the developed data-­driven models constructed from
practical data are in compliance with shipping domain knowledge is to explicitly
impose the constraints when developing the models. The following two sections in
this chapter introduce some initial thoughts on how to incorporate shipping domain
knowledge into the development of data-­driven models used to deal with a specific
problem in maritime transportation. To be specific, section 13.1 discusses how to
consider feature monotonicity into a tree-­based model developed to predict ship
risk, and section 13.2 discusses how to jointly consider feature monotonicity and
convexity into ship fuel consumption rate prediction.

13.1 Considering feature monotonicity in ship risk prediction

The content covered in this section is mainly from a paper published by Yan et al. in
2021 [1]. That paper aims to assist the Marine Department of Hong Kong to select
ships of higher risk among all the foreign visiting ships for inspection by using
state-­of-­the-­art ML (machine learning) models with shipping domain knowledge
considered. An eXtreme Gradient Boosting (XGBoost) model is first developed to
predict the number of deficiencies each foreign visiting ship has. In particular, the
XGBoost model takes domain knowledge regarding ship flag performance, RO (rec-
ognized organization) performance, and company performance in PSC into account,
i.e., given two ships with all the other conditions equal (e.g. with the same age,
of the same type, with the same last inspection time, and inspection results), the
ship that has worse flag performance, or RO performance, or company performance,
should be predicted to have a larger number of deficiencies and thus, should have
a high priority/probability to be inspected. The ship risk prediction results are then
input to a PSC officer scheduling model to help the maritime authorities to allocate
scarce inspection resources such that as much ship noncompliance as possible can
be detected. As only the first part of this study is related to ML model development,
we only cover that part in this section. Our aim is to discuss how to incorporate
domain knowledge into the model construction process of an XGBoost model using
more plain words, so as to make readers with a weaker background in data analytics
easier to understand.

13.1.1 Introduction of monotonicity in the ship risk prediction

problem
The monotonicity property of three features: ship flag performance, RO perfor-
mance, and company performance, which are also used by the Tokyo MoU to deter-
mine ship risk profile (SRP) in the New Inspection Regime (NIR) since January 1,
2014, is considered in this section. The SRP in NIR considers several ship-­related
parameters including type of ship, age of the ship, ship flag performance, ship RO
performance, ship company performance, deficiency conditions, and detention con-
ditions in the previous 36 months [2]. Different weighting points are attached to
different states of the parameters, and the final ship risk is determined by the total
weighting points. Especially, ship flag performance is established on an annual basis
 Incorporating shipping domain knowledge into data-­driven models  181

taking into account the performance of the ships under a certain flag state regarding
both inspection and detention conditions over the preceding three calendar years.
The flag performance is given by a black-­gray-­white list published by the Tokyo
MoU (Memorandum of Understanding) each year in its annual report. That is, ship
flag performance from the best to the worse is white, gray, and black. If the ships
under a flag state are inspected less than 30 times by all the member states of the
Tokyo MoU, the flag’s performance is set to “undefined.” An RO is an authorized
(and recognized) organization by a flag administration to carry out inspections and
surveys on its behalf on the ships under its registration. The performance of RO can
be high, medium, low, and very low from the best to the worst, which is determined
by its ships’ inspection and detention history over the preceding 3 calendar years,
and is published by the Tokyo MoU in its annual report. Similar to ship flag per-
formance, if the number of inspections of the ships in an RO is less than 60 times
over the last 3 years, the performance of this RO is “undefined.” The calculation of
ship company performance is quite different from the above two indicators. Ship
company here refers to a ship’s international safety management company. Its per-
formance considers the detention and deficiency conditions of the ships in its fleet,
which is updated on a daily basis for a running 36-­months period. Unlike ship flag
and RO performance, there is no lower limit on the number of inspections to quan-
tify the performance of a company over the last three years. For a company with no
inspection in the previous three years, it will be given two weighting points. The
company’s performance is determined by its detention index and deficiency index,
and it also has four states, namely high, medium, low, and very low.
The above three risk factors are considered in the NIR in the following way.
The criteria for a ship to be an LRS (low-­risk ship) are that its flag is white and is
of the IMO (International Maritime Organization) audit, its RO is recognized by at
least one member authority of the Tokyo MoU and it should have high performance,
and its company should have high performance. In addition, if a ship’s flag is on
the black list or a ship’s RO has low or very low performance, 1 weighting point is
attached to the ship, respectively. If a ship’s company performance is low or very
low, or there is no inspection on the ships belonging to the company in the previ-
ous 36 months, 2 weighting points are attached to the ship. It should be noted that
these three indicators considered in the NIR are subjective, as they are quantified
using criteria established by expert knowledge regarding the division of the states,
the ways to generate the states, and the weighting points attached to the states. The
common sense or domain knowledge, regarding these three risk factors is that for
two ships with the other features identical, the ship that has worse flag/RO/company
performance is expected to perform worse in the current PSC inspection, i.e., to have
a larger number of deficiencies and/or a higher probability of detention. This domain
knowledge should also be followed by data-­driven models developed for ship risk
prediction for the following reasons. It is obvious that ship flag, RO, and company
play an important role in ship management, operation, and maintenance, and hence
their performance is taken into account by the widely adopted NIR for ship risk
calculation and selection for PSC inspection. In return, their ships’ performance in
PSC inspections would influence their performance evaluated by the corresponding
182  Machine learning and data analytics for maritime studies

MoUs and thus their reputation. Therefore, it is justifiable to conclude that given all
other conditions being equal, a ship should be estimated to have worse performance
if its flag, RO, or company gets worse. Only by following such a rule can a ship risk
prediction model be regarded to be in compliance with shipping domain knowl-
edge, and thus achieve fair prediction. Theoretically, such domain knowledge can be
learned from practical inspection records. Nevertheless, as actual data are featured
with noise and error, it cannot be guaranteed that such domain knowledge can be
preserved in the data-­driven models constructed. In other words, if no constraint is
imposed in the model training process, such property may not be fully followed,
leading to unexpected prediction results.
In Yan et al. [1], the problem of ship deficiency number prediction in PSC to assist in
high-­risk ship selection for PSC inspection is addressed. The dataset contains 1974 initial
inspection records at the Hong Kong port, and the input to the ship risk prediction model
includes several ship-­related features as well as ship historical inspection features. To be
more specific, 14 features are considered, namely ship age, GT (gross tonnage), length,
depth, beam, type, flag performance, RO performance, company performance, the last
inspection date in the Tokyo MoU, the number of deficiencies in the last inspection in
the Tokyo MoU, the total number of detentions in all historical PSC inspections, ship flag
change times, and whether the ship has a casualty in the last 5 years. Inspection records
with ship flag and RO performance undefined are first deleted from the initial dataset, and
1926 records remain in the dataset. Then, the states of ship flag performance are encoded
by setting white to 1, gray to 2, and black to 3, and the states of ship RO and company
performances are encoded by setting high to 1, medium to 2, low to 3, and very low to 4.
Therefore, the domain knowledge regarding ship flag/RO/company performance and the
ship deficiency number can be viewed as a monotonic increasing relationship between
them: as the value of ship flag/RO/company performance gets large (i.e. from good to
bad), the ship deficiency number also gets large.

13.1.2 Integration of monotonic constraint into XGBoost

After data processing, an XGBoost model possessing such monotonicity property
is developed for ship deficiency number prediction. XGBoost is an implementation
of a gradient boosting machine, whose principle is to construct new weak learners
that are maximally correlated with the negative gradient of the loss function associ-
ated with the whole ensemble in the current state. We do not introduce how to con-
struct an XGBoost model in detail here, and we refer the readers to Reference [3],
which originally proposes the XGBoost model, and Reference [1], which develops
an XGBoost model for ship deficiency number prediction in PSC with details given
to model construction process, for more information. In this section, we only discuss
how to combine feature monotonicity characteristics into the development process of
an XGBoost model.
The base learner of an XGBoost takes a tree structure constructed by the CART
algorithm, and the learning objective of a base learner is to minimize the balance
between model prediction error (i.e. the difference between the predicted and actual
numbers of ship deficiencies, where half of the MSE (mean squared error) is often
 Incorporating shipping domain knowledge into data-­driven models  183

used) and model complexity (evaluated by the number of leaf nodes contained in a
decision tree). Recall that as introduced in Chapter 9, a classic DT (decision tree)
does not possess the monotonicity property. Therefore, we will introduce in detail
how to enforce a monotonic constraint to one feature (which can be ship flag perfor-
mance, RO performance, or company performance) regarding the number of defi-
ciencies predicted.
We first assume that feature monotonicity is imposed on one feature. As fea-
ture monotonicity works in the context that all the other features are equal except
for the feature imposed by the monotonicity property, without loss of general-
ity, we assume that the feature imposed by such property is ship flag performance.
Suppose that we consider a total of ‍m‍ features for ship deficiency number predic-
tion where the number of deficiencies of ship ‍s‍ is denoted by ys, and the feature vec-
0
tor is denoted by ‍x = (x1 , ..., xm , ..., xm ).‍ We further denote ship flag performance
by ‍mN ‍ where the monotonic increasing constraint is imposed. This means that for
two ships s1 and s2, the monotonicity property of ship flag performance works when
m0 m0 0 N + 1, ..., m,‍ and given ‍xmsN1 < xmsN2,‍ we should have the pre-
‍xs1 = xs2 , m = 1, ..., m
N  1, m
dicted numbers of ship deficiencies ‍yO s1  yO s2.‍ As the property of feature monotonicity
works in examples with all features being equal except for the feature enforced by the
monotonicity property, the tree can be simplified by only containing the splits using the
monotonic feature, as using all other features (which are identical to all the examples
considered) for node splitting will always lead these examples to the same tree nodes
and thus have the same output.
The structure of a simplified tree in an XGBoost model which only takes the
monotonic feature into account is shown in Figure 13.1. It should be emphasized

Figure 13.1  An illustration of imposing monotonicity constraint to a tree in an

XGBoost model
184  Machine learning and data analytics for maritime studies

that only the values of the feature imposed by the monotonically increasing con-
straint will be used for node splitting, as this constraint works in the context that all
the other features are identical for the related ships. Similar to a traditional CART
regression tree, the tree starts to split from the root node at the top which contains the
whole training set, and the output of this node is denoted by O ‍ T ‍. Suppose further that
‍a ‍is used as the threshold of feature ‍xmN ‍to split the root node into the left child node
‍L‍and right child node ‍ R‍, i.e., for examples with feature value no larger than ‍a ‍, they
are split to the left child node whose output is O ‍ L‍, and the other examples are split to
the right child node whose output is O ‍ R‍. To guarantee the monotonically increasing
constraint put on feature ‍xmN ‍, constraint O ‍ L  OR‍ is imposed to the outputs of the
child nodes of the root node. If this constraint cannot be satisfied when using ‍a ‍ as
the threshold for splitting, it will not be used to split the root node. Furthermore, if
no candidate value of feature ‍xmN ‍can satisfy the constraint, the node will not be split.
The above two rules apply for all nodes in a tree. Therefore, it can be guaranteed that
if the root node can be split, we have O ‍ L  OR‍. Similarly, constraint O ‍ LL  OLR‍ is
imposed to node ‍ L‍ and O ‍ RR  ORL‍is imposed to node ‍ R‍, such that the monotonically
increasing constraint can be guaranteed for nodes ‍ L‍ and ‍ R‍, and we can expect that
node ‍L‍and node ‍ R‍are either not split, or have O ‍ LL  OLR‍and O ‍ RR  ORL‍in the child
nodes, respectively. Furthermore, as the output of node ‍ LR‍should be no more than
the output of node ‍ R‍, an upper bound is enforced to the output of node ‍ LR‍, i.e., O ‍ LR‍,
OL +OR
as O ‍ LR  mean(O L , O R ) = 2  O R .
‍ Similarly, a lower bound is enforced to the
OL +OR
output of node ‍ RL‍, O‍ RL‍, as O
‍ RL  mean(O L , O R ) = 2  O L .
‍ Therefore, the out-
puts of the nodes in this layer satisfy that O ‍ LL  OLR  mean(OL , OR )  ORL  ORR.‍
As the tree is split in a recursive manner, the monotonicity increasing property of
the whole tree and thus on feature ‍xmN ‍ can be guaranteed. As the gradient boosting
machine is in an additive manner, the monotonicity property of the whole XGBoost
model can be guaranteed, and so are the final predicted ship deficiency numbers.
An XGBoost model imposed by monotonic constraint is denoted by monotonic
XGBoost by Yan et al. [1]. Yan et al. further compare the monotonic XGBoost
model with other popular and state-­of-­the-­art regression models, including standard
XGBoost, monotonic LightGBM, DT, RF (random forest), GBRT (gradient boost-
ing decision tree), LASSO (least absolute shrinkage and selection operator) regres-
sion, ridge regression, and SVR (support vector regression), and conclude that the
prediction performance of the monotonic XGBoost is the best among all the models
for comparison regarding both MSE and MAE. This complies with the comment
that model prediction performance can be improved if reasonable monotonic con-
straints on certain features are imposed, showing that such constrained models can
generalize better [4–6].

13.2 Integration of convex and monotonic constraints into ANN

(artifical neural network)

This section aims to address the ship fuel consumption rate prediction problem
while taking shipping domain knowledge regarding ship sailing speed and fuel
 Incorporating shipping domain knowledge into data-­driven models  185

consumption into account. As the shipping industry is mainly powered by heavy fuel
oil, a large environmental footprint has been created by shipping activities due to the
emissions of greenhouse gases (GHGs) and air pollutants. According to the report
of the fourth IMO GHG emission, shipping was responsible for 2.76% of global
anthropogenic GHG emissions in 2012, and the proportion increased to 2.89% in
2018 [7]. In recent years, the sustainability of shipping has become a public con-
cern, and various emission control measures and regulations have been imposed on
ships involving both national and international transportation activities to reduce the
adverse impacts on the environment. Meanwhile, the cyclical downturn of the world
economy and high bunker prices make it necessary and urgent for ship owners and
managers to operate their ships in a more cost-­effective way while still accomplish-
ing the global trade amount. In ship’s daily operations, bunker fuel costs account
for a large proportion of vessel operating costs. For large vessels such as container
ships, when the bunker fuel price is about 500 USD per ton, the bunker cost can con-
tribute to about 75% of the operating cost [8, 9]. Under these conditions, shipping
companies are constantly making hard efforts to optimize ship energy efficiency, i.e.,
to use as little fuel as possible to accomplish the required transport tasks.
Various factors can influence ship energy efficiency due to the complexity of
vessel engine systems and the surrounding sea and weather conditions in a voyage.
According to Yan et al. [10], common features that can influence ship fuel con-
sumption rates can be divided into four categories: ship mechanical features, ship
operational features, ship maintenance features, and sea and weather conditions.
Especially, ship mechanical data include ship dimension features (e.g. length, beam,
and gross tonnage) and power system features (e.g. engine parameters and designed
speed). Ship operational features include ship voyage and sailing behavior informa-
tion like sailing speed over ground, sailing speed through water, type of fuel used,
fuel density, and temperature, etc., as well as ship mechanical conditions while oper-
ating, such as the conditions of propeller pitch, rudder angle, main engine load and
working hours, hull and propeller fouling conditions, and wetted surface area. Ship
maintenance features mainly include ship dry docking records. Sea and weather
condition data mainly refer to the sea states and weather information along the voy-
age, where the sea states include sea depth, sea water temperature and density, wave,
swell, and current conditions, etc., while weather information includes the direction
and value of wind, air density, and temperature, etc.
Among all the influencing factors on ship fuel consumption rates, ship sailing
speed is widely believed to be the most significant determinant of ship fuel con-
sumption. In many existing studies, a ship’s fuel consumption rate at sea is usually
treated as proportional to its sailing speed to a power of ‍ ˛‍, where ‍ ˛‍is empirically
shown to be from 1.452 to 4.8 as summarized in Reference [10]. This shows that
ship hourly fuel consumption takes a monotonically increasing and convex relation-
ship with ship sailing speed. In other words, the fuel consumption rate is higher
when a ship’s speed is higher when all other things being equal, and the ship fuel
consumption rate increases faster as the sailing speed increases, as the ship fuel con-
sumption rate is approximately proportional to speed to a power larger than 1. The
domain knowledge regarding the relationship between ship sailing speed and ship
186  Machine learning and data analytics for maritime studies

Figure 13.2   An illustration of a standard ANN model

fuel consumption rate should be taken into account when developing data-­driven
models for ship fuel consumption prediction. This section discusses the idea of
incorporating such domain knowledge into the development of ANN models.
Traditional ANNs are introduced in chapter 8 of this book. Suppose that we have
a traditional ANN model shown in Figure 13.2 on hand, where the input layer with
‍m + 1‍ neurons is used to receive the input of ‍m‍ features, which can be the average
sailing speed, sailing condition, draft, air density, sea water temperature, and wave
and wind conditions, and the ‍(m + 1)th ‍ neuron is the bias term. The hidden layer has
‍K + 1‍neurons, where the (‍ K + 1‍)th neuron is the bias. The output layer gives the pre-
dicted ship fuel consumption rate and is denoted by ‍yO .‍ Without sacrificing generality,
we denote by node x1 in the input layer ship sailing speed, which is imposed by both
convex and monotonically increasing constraints, while the other nodes are free from
constraints imposed. To preserve such domain knowledge, we introduce a special
type of neuron in the hidden layer named convex and monotonic neuron, which has
convex and monotonically increasing activation functions and non-­negative weights
connected to them. These weights include those from the input layer and those to the
output layer. In addition, it is also required that node x1 representing the feature of ship
sailing speed in the input layer only connects to convex and monotonic neurons in the
hidden layer. The ANN model imposed by the above features based on Figure 13.2 is
denoted by ANN-­DK and is shown in Figure 13.3.
Figure 13.3 shows that node x1 representing ship sailing speed is connected
with nodes b1 to ‍bK1‍ in the hidden layer marked in orange that has monotonically
 Incorporating shipping domain knowledge into data-­driven models  187

increasing and convex activation functions using non-­ negative weights. These
nodes are also connected with the output layer using non-­negative weights. It is
also noted that node x1 is only connected to the aforementioned nodes with mono-
tonically increasing and convex activation functions while there is no connection
between node x1 and the other nodes without constraints imposed in the hidden
layer. Moreover, to guarantee the non-­negativity of the weights, if the weight values
updated in each round of training is positive or zero, there is no further action; oth-
erwise, their values are changed to zero. For the other nodes in the input layer, i.e.,
nodes x2 to ‍xm+1‍without constraints imposed, they are connected to nodes b1 to ‍bK1‍
as well as the other normal nodes in the hidden layer marked in yellow. The follow-
ing Theorems 13.1 and 13.2 guarantee that the ANN-­DK model is monotonically
increasing and convex in x1, and thus the predicted output is in compliance with
shipping domain knowledge.
Theorem 13.1 The following relation holds, which guarantees the monotonicity
property of x1 to the predicted fuel consumption rates:
@Oy
 0. (13.1)
‍ @x1 ‍
Proof:
@Oy
PK @Oy zk tk
= k=1 zk tk x1
P
@x1
K
= v f (tk )w1k
k=1 k
0

P K1 PK (13.2)
= v f (tk )w1k +
k=1 k
0
k=K1 +1 vk f0 (tk )w1k
P K1
‍ = v f (tk )w1k  0,
k=1 k
0
‍

Figure 13.3   An illustration of a domain knowledge informed ANN model

188  Machine learning and data analytics for maritime studies

where tk is the weighted sum of the inputs to node bk in the hidden layer. As the
activation functions are monotonically increasing in the input to nodes b1 to ‍bK1‍
in the hidden layer, ‍f0 (tk )‍ is positive. Furthermore, considering that weights vk and
PK1 0
‍w1k , k = 1, ..., K1‍ are non-­negative, it can be guaranteed that ‍ k=1 vk f (tk )w1k  0.‍
Meanwhile, as there is no connection between node x1 and nodes ‍bK1 +1‍ to ‍bK ‍, we
P @Oy
have ‍ Kk=K1 +1 vk f0 (tk )w1k = 0.‍ Therefore, @x is non-­negative, showing that x1 is
‍ 1‍
monotonically increasing in the predicted fuel consumption rate by the ANN-­DK
model.
Theorem 13.2 The following relation holds, which guarantees the convexity of
the predicted fuel consumption rate in x1:

@2 yO
 0. (13.3)
‍ @x21 ‍
Proof:
PK
@2 yO @ k=1 vk f
0 (t
k )w1k
=
@x21 @x1
PK @tk
00 (t
= k=1 vk w1k f k)
@x1 (13.4)
PK1 00 @tk PK @tk
= k=1 vk w1k f (tk ) + k=K1 +1 vk w1k f00 (tk )
@x1 @x1
PK1 2 00 @tk
= k=1 vk w1k f (tk )  0.
‍ @x1 ‍

The above inequality holds because the activation functions of nodes b1 to ‍bK1‍ are
convex, there are no connections between node x1 and nodes ‍bK1 +1‍to ‍bK ‍in the hidden
layer, and weight ‍w1k ‍is non-­negative.
Therefore, it can be safely concluded that as long as Theorems 13.1 and 13.2 can
be satisfied in an ANN model developed for ship fuel consumption prediction, the
convex and monotonically increasing relationship between ship sailing speed and
hourly fuel consumption rate can be preserved. Furthermore, such constraints can
be imposed on more features. For example, when the direction of wind/swell is the
same as the direction of a ship’s heading, a larger wind force/swell would generally
decrease the hourly fuel consumption rate. To consider such a monotonic relation-
ship, another new type of neuron with a monotonic activation function (which is not
necessarily convex) can be introduced in the hidden layer, which is also connected
by non-­negative weights with the corresponding neurons imposed by the mono-
tonic constraint in the input layer and the neurons in the output layer. ANNs with
the above properties can also be applied to address other problems where similar
domain knowledge should be preserved both within and outside of the area of mari-
time transportation.
 Incorporating shipping domain knowledge into data-­driven models  189

References

[1] Yan R., Wang S., Cao J., Sun D. ‘Shipping domain knowledge informed pre-
diction and optimization in port state control’. Transportation Research Part
B. 2021;149:52–78.
[2] Information sheet on the new inspection regime (NIR) [online]. Tokyo:
Tokyo MoU. 2022. Available from http://www.tokyo-mou.org/doc/NIR-
information%20sheet-r.pdf
[3] Chen T., Guestrin C. ‘XGBoost: a scalable tree boosting system’. Proceedings
of the 22nd ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining; 2016. pp. 785–94.
[4] Sill J. ‘Monotonic networks’ in Advances in Neural Information Processing
Systems; 1997.
[5] Daniels H., Velikova M. ‘Monotone and partially monotone neural networks’.
IEEE Transactions on Neural Networks. 2010;21(6):906–17.
[6] Pei S., Hu Q., Chen C. ‘Multivariate decision trees with monotonicity con-
straints’. Knowledge-­Based Systems. 2016;112:14–25.
[7] Fourth IMO GHG study 2020 – final report [online]. London: IMO
Publications. 2020 Aug 25. Available from https://www.imo.org/en/OurWork/​
Environment/Pages/Fourth-IMO-Greenhouse-Gas-Study-2020.aspx
[8] Ronen D. ‘The effect of oil price on containership speed and fleet size’.
Journal of the Operational Research Society. 2011;62(1):211–16.
[9] Wang S., Meng Q., Liu Z. ‘Bunker consumption optimization methods in
shipping: a critical review and extensions’. Transportation Research Part E.
2013;53:49–62.
[10] Yan R., Wang S., Psaraftis H.N. ‘Data analytics for fuel consumption man-
agement in maritime transportation: status and perspectives’. Transportation
Research Part E. 2021;155:102489.
This page intentionally left blank
 Chapter 14
Explanation of black-­box ML models in
maritime transport

In addition to model fairness, another serious drawback of using ML (machine

learning) models to address practical problems in maritime transport is related to the
black-­box property of most ML models: despite their success in many real-­world
applications to not only the maritime industry but also other industries including
defense, medicine, finance, and law, thanks to their high prediction performance,
they are opaque in terms of explainability, which makes users and even developers
difficult to understand, trust, and manage such powerful AI (artificial intelligence)
applications. Especially, when decisions derived from black-­box systems based on
ML models affect human’s life, safety, and the environment, there is a more urgent
need for explaining and understanding how such decisions are furnished by AI
methods. Especially, in the relatively traditional and conservative maritime indus-
try, decision-­makers and stakeholders are more likely to be reticent to adopt deci-
sion support tools powered by new technologies such as AI which they can hardly
interpret, control, and thus trust. This chapter aims to first introduce the necessity of
explaining black-­box ML models in the maritime industry, especially in the case of
PSC (port state control). Then, popular methods to achieve black-­box model expla-
nation are introduced.

14.1 Necessity of black-box ML model explanation in the

maritime industry

14.1.1 What is the explanation for ML models

Before discussing the details of providing explanations for black-­box ML models,
we first distinguish two similar yet different terms: interpretability and explanation.
We have to say that there is no unified definition of either term in the existing litera-
ture within the context of ML or AI, and they may have different meanings in differ-
ent scenarios. We borrow the definition of interpretability given by Doshi-­Velez and
Kim [1], which says that
Definition 14.1: Interpret means to explain or to present the understandable
terms.
192  Machine learning and data analytics for maritime studies

In the domain of ML, the authors further define interpretability as the ability to
explain or present in understandable terms to a human. Then, according to Arrieta et
al. (2020) [2], for ML models, interpretability is “a passive characteristic of a model
referring to the level at which a given model makes sense for a human observer.”
The authors further explain that this characteristic can be interpreted as model trans-
parency. This shows that model interpretability is an inherent property of an ML
model, indicating that the ML model itself processes the ability to make the develop-
ers and users understand its reasoning process and the predictions generated. Typical
ML models falling in this category include rule-­based learning and reasoning, linear
regression (LR), decision tree (DT), and kNN (k-­nearest neighbor). Details of these
ML models will be covered in the next sections. In contrast, explainability is “an
active characteristic of a model, denoting any action or procedure taken by a model
with the intent of clarifying or detailing its internal functions.” Therefore, model
explainability can be understood as explaining the ML model of black-­box nature
developed using other external techniques, where such techniques can be local (i.e.
for the prediction of a single example) or global (i.e. for the overall performance of
the whole model), and model-­specific (which are limited to specific model classes)
or model-­agnostic (which can be used to any ML models). A detailed introduction
of the related methods will be given in the next sections.
We say that most of the ML models are of black-­box nature because we do not
know how the features fed into the model are processed and used to give the final
output. According to Doshi-­Velez and Kim [1], the black-­box nature of ML models
is not a big deal when (1) there are no significant consequences for unacceptable
results given by an ML model; or (2) the problem is sufficiently well studied and val-
idated in practice, and thus, the users trust the decisions recommended by the system
even if it is imperfect, i.e., the prediction results are not always of 100% accuracy.
Otherwise, an ML model explanation should be given to address the incomplete-
ness in the problem formalization, which means that getting to know what if the
prediction is not enough for some problems, and it is also expected that why such
prediction is made should be explained [3]. Unfortunately, neither condition is held
in the maritime industry. The main reason is that this industry involves several het-
erogeneous and conservative stakeholders and decision-­makers, who make strate-
gic, tactical, and operational decisions heavily dependent on long-­term experience
instead of emerging data-­driven models, especially ML models. Consequently, rec-
ommendations generated by ML models in black-­box nature without convincing
explainability provided can be seldom acceptable to them, even if in many cases the
recommendations given by these models can be much more efficient and reasonable
than transparent but naive rules or expert systems.
One example in maritime transportation is ship selection decisions in PSC made
on each workday at port authorities: although there are various high-­performing data-­
driven models developed for ship risk prediction or high-­risk ship recommendation
in existing studies, which have also been empirically shown to be more efficient than
the rule-­based SRP (ship risk profile) ship selection method adopted by most MoUs
(Memorandum of Understandings), these data-­driven models are seldom adopted by
port authorities at the moment. Meanwhile, for ship risk prediction problem in PSC,
 Explanation of black-­box ML models in maritime transport  193

the first condition mentioned above is not satisfied, because ship selection decision is
vital for both port authorities and ship operators. For port authorities, the number of
foreign visiting ships is large each day, while the available inspection resources can
be scarce. Moreover, it can be seen from the annual reports of the MoUs that only
some of the inspected ships are with deficiency/deficiencies identified, while only a
small proportion of them are actually detained, i.e., with serious deficiency/deficien-
cies found. Therefore, only a small proportion of all the foreign visiting ships can be
and should be inspected. It is thus crucial to allocate the limited inspection resources
to the ships with the highest risks (i.e. with the worst performance) in PSC inspec-
tion to improve the efficiency of inspection. From the perspective of ship operators,
too frequent and unnecessary inspections would delay the shipping schedule, and
thus, reduce the efficiency of the fast turnover of the shipping logistics system and
cause financial loss. Besides, if too few substandard ships are inspected by the port
states, ship operators may lack the motivation to intensively maintain their ships to
be in good condition. On the contrary, if too many qualified ships are inspected by
a certain port, ship operators may turn to other destinations with looser inspection
strategies, and such behavior is referred to as “port shop,” which may affect overall
vessel quality and ports’ reputation. Besides, the second condition mentioned above
is also not satisfied in ship selection for inspection by PSC, because although this
topic is widely studied in existing literature where several data-­driven models have
been developed to assist port authorities in high-­risk ship selection, the proposed
models are not applied by the ports, and thus, the results have not yet been fully
validated. Consequently, there is still a long way to go to make the users trust the
prediction results given by the data-­driven models in PSC.
In addition to fulfilling the needs of model explanation according to Doshi-­Velez
and Kim [1], providing explanations to black-­box models developed to address
practical problems in maritime transport can bring the following advantages:

1. Trust: conservative stakeholders and decision-­makers in the traditional mari-

time industry are reluctant to accept and use recommendations given by data-­
driven models based on empirical data as these models are not as tractable and
controllable as the expert knowledge systems currently adopted. Therefore, to
gain their trust in the newly proposed data-­driven models and thus, promote
the use of these models in practical decision-­making, explanations are needed
to help them understand and verify the data-­driven models’ internal schemes,
working processes, and strengths as well as weaknesses.
2. Transferability: since explanation can shed light on understanding and imple-
mentation of the black-­box models developed, it can help the users to know
how well the data-­driven prediction models generalize in different scenarios and
thus, can promote the use of the data-­driven models.
3. Fairness: as explanations reveal how a black-­box data-­driven model works and
why a prediction is given, model fairness can be verified to ensure that the pre-
dicted results do not discriminate against particular stakeholders. Especially, as
decisions in the maritime industry would influence a variety of stakeholders,
such as ship owners, operators, management companies, port authorities, and
194  Machine learning and data analytics for maritime studies

shipping service providers, model fairness is a key to guarantee that the predic-
tions are reasonable and thus, verify the recommendations given by the black-­
box models to be compliant with ethical standards and common beliefs.
4. Extensibility: explanations given for black-­box data-­driven models can help the
developers to further adjust the settings and hyperparameters of the ML models
more effectively, so as to improve their prediction accuracy. They can also help
the model users generate general rules and extract new knowledge from the
massive data for future decision-­making.

14.1.2 Propose and evaluate explanations for black-box ML models

Then, in order to provide a proper explanation of a black-­box model, what ele-
ments should be considered when explaining? Basically, two questions need to be
answered when explaining black-­box ML models: how and why [4]. The “how”
question refers to how the black-­box model works, i.e., its internal working mecha-
nism, and the “why” question refers to why a certain output is predicted for a certain
example. However, it is also pointed out in existing literature that different explana-
tions are expected to be given to distinct behaviors, distinct problems, and distinct
users in Vilone and Longo [4]. In other words, black-­box model explanations should
consider the following four elements: the user, i.e., the receiver of explanations;
the goals, i.e., what questions need to be answered by the explanations; the con-
tents, what information is expected to be obtained from the explanations; and the
language, i.e., what terminology should be used when presenting the explanations.
These four elements are correlated with each other. Details and relations between
them are summarized in Table 14.1.
After presenting an explanation to a black-­box model considering the above
four elements, a natural question then is which explanation is good? In other words,
what are the criteria to evaluate the quality of explanations given to a black-­box
model. Actually, it is still not that clear what is the most effective way to structure
explanations according to Vilone and Longo [4], and how to select and evaluate
model explanation methods for a particular problem and audience as mentioned in
Murdoch et al. [5]. This indicates that there might be a lack of unified ways to evalu-
ate the explanations given to a specific problem, as problems themselves can vary a
lot from each other. The evaluation process is quite different in model explanation
and model construction using supervised ML techniques as introduced in the earlier
chapters of this book, where there are unified and widely recognized metrics to
quantitatively evaluate model prediction performance. For model explanation evalu-
ation, we introduce a framework called PDR, which is short for predictive accuracy,
descriptive accuracy, and relevancy as proposed in Murdoch et al. [5] for selecting
and evaluating the most suitable explanations given to a black-­box prediction model
for a specific problem. Details of the three points are provided as follows:

•• Predictive accuracy: predictive accuracy tends to evaluate the quality of the

black-­
box prediction model itself, as only explanations given to accurate
 Explanation of black-­box ML models in maritime transport  195

Table 14.1   Four elements of black-­box model explanation

User Goals Contents Language

Researchers and Understand how the Comprehensive Can be sophisticated
developers entire data-­driven and technical and full of
model works to explanations. terminologies
verify its fairness, to maintain
transferability, professionalism,
and extensibility, and might be hard
and to improve to understand by
its overall people who are not
performance. the researchers or
developers of the
model.
Model users Make sure that the Reasoning process The terminologies
current model of all components used should be
can meet the of the model and within the context
requirements its application of model practical
and purposes scenarios. use and avoid too
according to their professional terms.
expectations.
People affected by Get to know the The general Plain words without
the model but rationale behind model working professional
not direct users the black-­box mechanism and terminologies.
model without decision process.
being familiar
with all of
the model’s
components.

prediction models are meaningful. This requires that the prediction models to
be explained should learn the underlying relationships in the data well and thus,
generalize well to unknown examples. In general, model predictive accuracy
is evaluated by the accuracy of the test set using proper metrics. Under the
context of predictive accuracy evaluation for a model explanation, one should
pay attention to the data used to check the predictive accuracy: the test set data
must resemble the interest of the model users. For example, to evaluate a high-­
risk ship selection model developed for the Hong Kong port using inspection
records there, it is not reasonable to evaluate the prediction model’s perfor-
mance using the inspection records at a port other than the Hong Kong port, say
the Port of Singapore. Moreover, sometimes using average prediction accuracy
might not be enough, as it is also expected that model performance should be
stable when there are data and model perturbations within a reasonable range.
This is because if the model has dramatic changes when there are slight changes
in the data and model, the explanations generated from the model might not be
trustworthy.
196  Machine learning and data analytics for maritime studies

•• Descriptive accuracy: when explaining the behavior of a black-­box predic-

tion model, a critical point is that the explanations should capture the relation-
ships between features and output learned by the prediction model precisely.
Otherwise, the explanations are meaningless. Nevertheless, it should be
acknowledged that accurately capturing such relations can be hard, especially
for complex black-­box models such as deep neural networks. Therefore, there
is a trade-­off between model accuracy and model explanation in practical use.
•• Relevancy: in addition to high accuracy, explanations should be relevant con-
sidering a particular audience for a chosen domain problem. As shown in
Table 14.1, different audiences, e.g., researchers and developers, model users,
and people affected by the model but not direct users, have different expecta-
tions of the explanations provided to them. Taking the prediction model of ship
risk for PSC inspection as an example, the first class of users, who are develop-
ers and maintenance staff of the ship selection system, might be interested in
the explanations of the inner working mechanism, i.e., how the model captures
the relationship between the features and the target, what are the intermediate
outputs within the model, and the rationale behind the predictions, so as to
evaluate and improve model performance. The second class of users, who are
usually officials in port authorities, might focus more on the ship risk predic-
tion process, especially why a ship is predicted to be of certain risk and why a
ship is predicted to be of higher risk than the other ship. Moreover, they may
also be interested in prediction fairness as discussed in chapter 13. The third
class of users, who are usually ship operators, managers, and owners, might be
interested in the performance of their ships predicted by the ship risk prediction
model so as to reduce the frequency of inspection of their ships and improve
their ships’ performance as well as model prediction fairness similar to the sec-
ond class of users.

One point that needs to be added is that there is a trade-­off between predictive
accuracy and descriptive accuracy: usually, a more complex black-­box prediction
model has higher predictive accuracy, but this means that its descriptive accuracy
might be low, as such a model might be hard to analyze. In contrast, a simple white-­
box prediction model can have a high descriptive accuracy as its behavior is easier
to be captured. However, its predictive accuracy might not be high enough.

14.2 Popular methods for black-box ML model explanation

14.2.1 Forms and types of explanations

Explanations of the predicted results of an ML model can have different forms,
where the main forms are summarized as follows:

1. Model internal: this type of explanation is usually given by interpretable models

intrinsically, such as the weights in LR models and the tree structure as well as
the node splitting criteria in DTs.
 Explanation of black-­box ML models in maritime transport  197

2. Feature importance: this method aims to show how important a feature is to the
prediction of the target, i.e., how much does the feature contribute to the predic-
tion of the target. It can be expressed as the importance value of a single feature,
or in a more complex form such as the pairwise feature interaction strengths.
3. Partial dependence: this method aims to show the marginal contribution of
the values of a feature on the predicted output, which is often called partial
dependence plots. As marginal contribution needs to be calculated, all the other
features other than the feature to be explored need to remain unchanged, and
different values of the feature to be explored should be traversed.
4. Data point: this method aims to use existent or newly created examples to show
the decision process of an ML model. One typical method on a single example
level is called counterfactual explanation, which aims to find similar examples
of the example to be explained by changing some of its features and observing
the changes in the output, so as to understand how decisions are made by the
prediction model.
5. Surrogate model: this method aims to find a surrogate model which is inher-
ently interpretable, and can mimic the behavior of the black-­box ML model to
be explained.

There are several criteria that can be followed to classify explanations given to
black-­box ML prediction models from different perspectives. One criterion is based
on model property as discussed above: an ML model is intrinsically explainable or
external explanation methods should be applied to explain the ML model. The first
class of explanation is called intrinsic explanation and the second class is called post
hoc explanation. Intrinsically explainable models are interpretable ML models with
simple structures and are self-­explainable. A typical example is LR which takes an
additive linear form, and the coefficients of the terms (i.e. features) can be viewed
as their weighting points (or the approximate importance scores if feature values
are normalized) while the final predicted target of a particular example is given by
the weighted sum of its feature values. Another typical example is DT, especially
shallow DTs, in which the reasoning process is shown by the (feature and value)
pairs selected for node splitting, and the predicted output of a particular example is
determined by the average/majority outputs of the examples in the node where this
example falls. In addition, feature importance can be generated from a DT: a feature
is regarded to be more important than the other if it can reduce node impurity to a
larger extent. It should be noted that the explanations given by an intrinsic model
can be from relatively limited aspects, and such an explanation sacrifices the accu-
racy of the prediction model. In contrast, post hoc explanation refers to the process
of developing an ML model of black-­box nature first and then using external expla-
nation methods to explain the model. If necessary, it can also be applied to explain
intrinsic models from wider perspectives.
Another two criteria consider the application scope of the explanation method.
One is that if the explanation is given to a single example, it is local. Alternatively, if
the explanation is given to the entire model, it is global. The other is that if the expla-
nation method can only be used to a specific model, e.g., artificial neural networks or
198  Machine learning and data analytics for maritime studies

DTs, the explanation method is model-­specific. On the contrary, if the explanation

method can be applied to all types of ML models, it is model-­agnostic.
In the following sections, we first use DT as an example to show how to gain
explanations from an intrinsic explanation model. Then, we introduce a post hoc,
local, and model-­agnostic explanation method to explain black-­box ML models
called Shapley additive explanations (SHAP).

14.2.2 Introduction of intrinsic explanation model using DT as an

example
We use a regression tree for ship deficiency number prediction as an example to
show how to generate explanations from a DT. The dataset used to construct the
regression DT is the same as that used in Example 1.2 in Chapter 9. To reduce the
complexity of the tree for ease of illustration, we set hyperparameters max_depth to
3 and min_samples_leaf to 1. The core code to generate a visualized DT using the
scikit-­learn API [6] and the PyDotPlus API [7] is as follows:

‍ ‍

The visualized DT is presented in Figure 14.1. The node at the top of the tree
shown in Figure 14.1 is the root node, which contains all the 1,500 ships, i.e., the
entire training set, used for tree training. The first line in the node shows the feature
and its value selected to split the node, i.e., feature gross tonnage (GT, whose unit
is 100 cubic feet) is selected to split the root node with a threshold value of 6,451.5.
This means that ships whose GT is no more than the threshold are split to their left
child node, where the process is shown by the “True” arrow under the root node,
meaning that splitting criteria “GT ‍ <=‍ 6,451.5” is satisfied in these ships. Other
ships, i.e., ships whose GT is more than 6,451.5 the unit of GT has been added to the
place where feature GT is mentioned, are split to the right child node, as shown by
the “False” arrow under the root node. The second line in the root node shows the
current MSE (mean squared error), i.e., impurity, of the node. The MSE of the root
node is 27.732. The third line shows the number of ships contained by the node. The
fourth line is the predicted ship deficiency number of this node, which is the aver-
age deficiency number of all the ships contained in the node. This value is 4.309 in
the root node, indicating that the average number of deficiencies of the ships in the
entire training set is 4.309.
 Explanation of black-­box ML models in maritime transport  199

Figure 14.1  Visualization of a regression DT for ship deficiency number

prediction

The tree is split in a recursive manner until the stopping criteria of max_depth
as 3 is met. Therefore, we can see that the depth of the tree is 3. The total number of
ships contained in all leaf nodes is equal to 1,500. It is also interesting to find that
although splittings in DTs aim to reduce impurity in subsequent nodes, which is
represented by MSE in this example, the MSE may not necessarily be reduced after
splitting a node. An example is the split of the root node to its left child node, where
the MSE increases from 27.732 to 81.853. This indicates that the left child node
might contain examples with extreme output values (i.e. very large ship deficiency
numbers) that have a big influence on the MSE of the node. This phenomenon is also
shown by the outputs of leaf nodes: for leaf nodes with the predicted ship deficiency
number close to the output of the root nodes, i.e., nodes 3, 5, and 6, their MSE values
are less than or even much less than the MSE of the root node. The outputs of the
other nodes are large, which is mainly because they contain ships with very large
deficiency number. Consequently, their MSE values are higher than the MSE of the
root node. This shows that extreme values are indeed contained in the training set,
and they can have a very large impact on a DT’s performance.
The above analysis shows the intrinsic explainable property of DT models: the
inner working mechanism of a DT, i.e., how the DT is split from the root node to
leaf nodes is shown by the (feature and value) pair; the examples contained, output
and error can be obtained for each node; and if there comes a new ship, how to clas-
sify the ship to a leaf node and decide its predicted deficiency number are all clear.
Model fairness can be verified based on these explanations. For example, from the
root node, it can be seen that a smaller ship tends to have a larger deficiency number
in PSC inspection, as it may not be managed that well compared to larger ships.
Moreover, from the second node on the right in tree depth 2, it can be seen that if
a ship has a larger deficiency number in the last initial inspection, it is predicted to
have a larger deficiency number by the DT for the current PSC inspection, which
complies with expert knowledge. Based on the splitting criteria and the MSE of
200  Machine learning and data analytics for maritime studies

each node, feature importance can be generated. As only three features are used
for node splitting in the tree in Figure 14.1, i.e., GT, company_performance, and
last_deficiency_no, only these features have importance scores, which are 6.8689,
1.7838, and 1.4505, respectively, showing that ship GT is the most important feature
determining ship deficiency number, followed by ship company performance and
the number of deficiencies in the last PSC initial inspection. For the detailed meth-
ods to calculate feature importance score in regression DT, readers are referred to
Reference 8 for more information. In addition, decision rules can be extracted from
the DT in Figure 14.1. For example, it can be concluded that if a ship has GT no
more than 6,451.5 and has unknown company performance, its deficiency number
is predicted to be 19.393, which is relatively high and deserves much attention from
the port authority. Another rule is that if a ship has a GT larger than 6,451.5 while
less than 24,294.5 and its last deficiency number is no larger than 3.5, its predicted
deficiency number is 4.122.

14.2.3 SHAP method
SHAP is proposed by Lundberg and Lee in 2017 [9], which is theoretically based
on Shapley value in game theory. SHAP is used to explain the output of the predic-
tion of an individual example (i.e. local) of an arbitrary machine learning model
(i.e. model-­agnostic) after it is constructed (i.e. post hoc). A base value, which is
the average output of the examples in the training set, is first calculated, and the
contribution of each feature value of the example to the predicted target is calcu-
lated. Then, the final output of the example is the sum of the base value and the
contributions of its feature values where such contribution is represented by mar-
ginal contribution, which can be viewed to be similar to a linear model taking a sum
format. SHAP is motivated by coalition game theory, where the “game” refers to the
prediction task of an example, the “players” are the features included in the predic-
tion model, and the “gain” is the difference between the final predicted target and the
base value, i.e., the sum of the contributions of the feature values. Given an example
0 0
‍xi ‍with ‍m‍features, the contribution of feature ‍m ,‍ ‍m = 1, ..., m,‍ e.g., ‍xi ‍is denoted by
i

‍ m0 ‍. Furthermore, given that the base value is y
‍‍ , which is the average output in the
training set, the predicted output of the example ‍xi ‍ denoted by ‍f(xi )‍ explained by
SHAP can be represented by the following additive linear function form:
X
m

f(xi ) =y + mi 0 . (14.1)

‍ m0 =1 ‍
The key point in SHAP is how to calculate the marginal contribution of the value
of each feature for the example to be explained, i.e., ‍mi 0 ,‍ e.g., i‍ ‍’s feature ‍m0 ‍. The
basic idea is to calculate the difference when considering and ignoring this feature,
which is thus the marginal contribution of the feature. Several efficient algorithms
are proposed to calculate the SHAP values of feature values efficiently. Readers are
referred to Lundberg and Lee [9] and Lundberg et al. [10] for more information,
and we do not cover the detailed algorithms in this chapter. Using SHAP to explain
the prediction of black-­box ML models has been implemented by the SHAP API
 Explanation of black-­box ML models in maritime transport  201

Table 14.2  Feature values and the corresponding SHAP values of the example
ship

Feature name Value SHAP value

GT 131,332 ‍ 1.584329‍

age 5 ‍ 0.317698‍

flag_performance 1 ‍ 0.038422‍

RO_performance 1 ‍ 0.059324‍

company_performance 2 ‍ 0.053150‍

last_inspection_time 10 ‍ 0.171819‍

last_deficiency_no 5 ‍0.355471‍
last_inspection_state 0 ‍ 0.038518‍

type_bulk_carrier 0 ‍ 0.058035‍

type_container_ship 1 ‍0.007297‍
type_general_cargo/multipurpose 0 ‍ 0.173660‍

type_other 0 ‍0.005028‍
type_passenger_ship 0 ‍ 0.002885‍

type_tanker 0 ‍0.047903‍

developed by Lundberg [11]. Following is an example to use the SHAP method to

explain the ship deficiency number prediction model based on random forest (RF)
in Example 9.4 of Chapter 9.
We first explain the predicted number of deficiencies of ships in the test set. The
following code is used to obtain the matrix of SHAP values on the test set based on
the RF model constructed by the training set and the base value in the training set:

‍ ‍

Then, we explain the prediction of a random ship in the test set. Feature values
and their corresponding SHAP values of the selected example ship are shown in
Table 14.2.
The sum of SHAP values of all features is  ‍ 2.082140‍. Given the base value as
4.320877, the final output of the deficiency number of this ship is 2.238737. The real
deficiency number of this ship is 3, and thus, the absolute prediction error of this
ship is 0.761263. Main contributors to the final predicted deficiency number can also
be shown in a visual manner in Figure 14.2.
202  Machine learning and data analytics for maritime studies

Figure 14.2  Visualization of the main contributors to the example ship’s

predicted deficiency number

As shown in Figure  14.2, the base value is 4.32 and the predicted value is
2.24, where both are shown directly on the axis. The main features of this ship that
decrease the predicted deficiency number are shown in blue, which are having GT
at 131,332 (being a relatively large ship), of ship age 5 (being a young ship), and
not being ship type general_cargo/multipurpose (not of a ship type associated with a
larger number of deficiencies). Meanwhile, the major feature that increases the pre-
dicted ship deficiency number is having 5 deficiencies in the last initial inspection
within the Tokyo MoU, which is larger than the average deficiency number over the
training set at 4.32. The figure further shows that the overall effects of the values
of all the features lead to the ship having 2.08 less deficiencies than the base value,
and thus, the final predicted deficiency number of the ship is ‍4.32  2.08 = 2.24‍ in
the current PSC inspection. Then, the decision process of this ship, i.e., why it is
predicted to have 2.24 deficiencies and what are the main contributors to the final
target, is clearly presented to the users.
We then use SHAP to generate more insights from the training set. We first use
a beeswarm plot with the y-­axis representing each feature and the x-­axis represent-
ing the features’ SHAP values with each dot representing a single ship in the training
set as shown in Figure 14.3. Feature values from low to high are shown by gradient
colors as shown on the right side of the figure. When multiple dots land at the same
‍x ‍position, they pile up to show the density. The core code to generate this figure is
given as follows:

‍ ‍

We then explore more on the feature values on the predicted deficiency number.

•• GT: for feature values shown by blue dots, i.e., for relatively small feature val-
ues, many of their SHAP values are larger than 0, indicating that being a smaller
ship is more likely to have the worse condition and thus, increases the number
of deficiencies. It is also noted that when the feature values are shown in blue,
they can also have negative SHAP values which would reduce a ship’s pre-
dicted deficiency number. This shows that the conditions of small ships might
 Explanation of black-­box ML models in maritime transport  203

Figure 14.3  Visualization of local explanation summary on the training set

using SHAP

be varied. In contrast, when the feature values are shown in red which refers to
large ships, their SHAP values are always less than ‍1‍, showing that large ships
are generally in good condition.
•• Last_deficiency_number: having a larger number of deficiencies in the last ini-
tial PSC inspection in the Tokyo MoU would no doubt increase the predicted
number of deficiencies in the current PSC inspection, and these ships are shown
by the red dots with positive SHAP values, as this indicates that the ship is
in a relatively bad condition. Bad performance in the last PSC inspection can
increase the final prediction by up to 6. Meanwhile, it is also interesting to find
that even if a ship performs relatively well in the last PSC inspection, it does
not necessarily mean that the ship could perform well in the current inspection,
and these ships are shown by the blue dots with positive SHAP values. This is
mainly because ships’ conditions can vary from time to time, especially when
the last inspection is long ago. In addition, different port states may have differ-
ent inspection strategies and different degrees of strictness. In general, it can be
seen that having less deficiency number in the last PSC inspection can reduce
the number of deficiencies identified in the current PSC inspection.
•• Age: the general condition of this feature is that older ships tend to have posi-
tive and slightly negative SHAP values while younger ships tend to have nega-
tive SHAP values. This is intuitive and easy to understand: the manufacturing
technology of older ships is not as exquisite as that of younger ships, and older
ships may also have more wear and tear as well as damage after a long period
of sailing.
204  Machine learning and data analytics for maritime studies

•• Company_performance: recall that the values of ship company performance are

encoded in this way: unknown->0, high->1, medium->2, low->3, and very low-
>4. Therefore, when the values of company performance are small as shown by
blue dots, they are usually associated with nearly zero or negative SHAP values
shown in a cluster, while some of them are also attached with positive SHAP
values shown by sporadic dots on the right-­hand side. When the values of this
feature are large as shown by red dots, the associated SHAP values can be either
positive or negative. This shows that generally, a ship’s worse performance
company would increase its predicted deficiencies and vice versa. Nevertheless,
it is not absolute as there are many ships within a company’s fleet and their
conditions can vary a lot.
•• RO_performance: recall that the values of ship RO (recognized organization)
performance are encoded in this way: unknown->0, high->1, and medium->2.
As a larger value of RO_performance indicates that the RO has worse perfor-
mance, and thus, it is with more positive SHAP values as shown by the red dots
on the right-­hand side. When RO_performance has a medium value, i.e., 2, the
associated SHAP value is near 0. When an RO’s performance is unknown, it is
also slightly positive.
•• Flag_performance: recall that the values of ship flag performance are encoded
in this way: unknown->0, white->1, gray->2, and black->3. The SHAP values
associated with ship flag performance show a clear tendency: when the value
is high as shown in red dots, i.e., when the flag is on the gray or black list, the
associated SHAP values are usually positive. Otherwise, when the flag is on
the white list or when it is unknown, the associated SHAP values are near 0 or
slightly negative.
•• Last_inspection_time: last_inspection_time refers to the period in the month
between the current and the last PSC inspections within the Tokyo MoU. The
basic trends of a ship’s last inspection time and its deficiency number in current
PSC inspection are that usually, a shorter last inspection time indicates a larger
number of deficiencies. This is mainly caused by the ship inspection priority
according to the SRP ship selection method: the length of the inspection time
window becomes longer from HRS (high-­risk ship) to SRS (standard-­risk ship)
and LRS (low-­risk ship). Therefore, ships in worse conditions would be more
frequently inspected, and thus, their last inspection time would be shorter.
•• Last_inspection_state: recall that the values of ships’ last inspection states are
encoded in this way: none->-1, no->0, yes->1. Most of the ships are not detained
in the last PSC inspection (1,376 out of 1,500), while 49 are detained and 75
have no previous PSC inspection. Therefore, large values of this feature are
scattered on the right-­hand side associated with positive SHAP values, while
smaller values might be bunched near 0 with heavy coverage. Therefore, most
of the dots for this feature is in red while seldom of them are in blue.
•• Ship type: this feature includes six sub-­features: type_bulk_carrier, type_con-
tainer_ship, type_general_cargo/multipurpose, type_passenger_ship, type_
tanker, and type_other. Especially, the risk tendency of ships belonging to
type_bulk_carrier, type_general_cargo/multipurpose, type_passenger_ship,
 Explanation of black-­box ML models in maritime transport  205

and type_tanker is obvious: bulk carrier, general cargo carrier/ship, multipur-

pose ship, and passenger ships are associated with large SHAP values, while
the tanker is associated with small SHAP values. In contrast, a container ship
is generally associated with non-­negative SHAP values, while a ship type other
than a container ship can be attached with slightly negative, zero, and slightly
positive SHAP values. Finally, for other ship types, as they contain various
types of ships, including but not limited to gas carrier, heavy load ship, spe-
cial purpose ship, livestock carrier, and tug, they could either increase the pre-
dicted ship deficiency number (i.e. associated with positive SHAP values) or
decrease the predicted ship deficiency number (i.e. associated with negative
SHAP values).

Based on the analysis of the beeswarm plot representing the relationship

between ship feature values and their corresponding SHAP values, how different
values of the features contribute to the final predicted ship deficiency number is
clear, and it can also be verified that the predictions given by the RF model are gen-
erally in compliance with shipping domain knowledge. The overall importance of
each feature on the whole training set can be further generated from the beeswarm
plot shown in Figure 14.3 by calculating the mean absolute SHAP values of the
feature among all the examples in the training set. This means that even if SHAP
values of a certain feature can be positive or negative, their effects would not be
offset. In this way, the local SHAP method can be extended to a global one as fea-
ture importance is at the global level. It can be anticipated that the larger a feature’s
mean absolute SHAP, the greater influence the feature has on the final prediction as
it can change the predicted target more. Feature importance shown in the bar chart
is presented in Figure 14.4.
It is shown in Figure 14.4 that ship GT is the major determinant of ship defi-
ciency number, followed by the last deficiency number and ship age. Intuitively, a
ship’s last inspection state, i.e., whether the ship is detained in the last PSC initial
inspection, should have a large influence on the predicted deficiency number in the
current PSC inspection. However, as ships with detention only constitute a very
small proportion of all the ships in the training set (49 detentions among all the
1,500 ships), its importance is reduced, and it only ranks 10th among all the features
regarding their importance. For ship management parties, ship company perfor-
mance has the largest impact, followed by RO performance and flag performance.
As ship type is divided into six sub-­types, it can be hard to evaluate the overall influ-
ence of ship type. Nevertheless, Figure 14.4 suggests that general cargo carrier/ship
and multipurpose ship, bulk carrier, and tanker can have a large influence on the
predicted ship deficiency number.
Based on the above illustration and analysis, it can be seen that the SHAP method
can be used to explain the predictions given to individual ships by calculating mar-
ginal contributions of the values of each feature for the example to be explained in a
local manner, to show the overall influence of feature values on the prediction target
over the whole data set in a local manner, and to generate feature importance in a
global manner. Based on the explanations, the model reasoning process and thus,
206  Machine learning and data analytics for maritime studies

Figure 14.4   Visualization of global feature importance using SHAP

fairness can be verified, so as to convince the users regarding the rationale behind
the black-­box model to improve model acceptance and practical applicability.

References

[1] Doshi-­Velez F., Kim B. ‘Towards a rigorous science of interpretable machine

learning’. ArXiv:170208608. 2017.
[2] Barredo Arrieta A., Díaz-­Rodríguez N., Del Ser J, et al. ‘Explainable artifi-
cial intelligence (XAI): concepts, taxonomies, opportunities and challenges
toward responsible AI’. Information Fusion. 2020;58:82–115.
[3] Molnar C. Interpretable machine learning. L ​ ulu.​com; 2020.
[4] Vilone G., Longo L. ‘Notions of explainability and evaluation approaches for
explainable artificial intelligence’. Information Fusion. 2021;76:89–106.
[5] Murdoch W.J., Singh C., Kumbier K., Abbasi-­Asl R., Yu B. ‘Definitions,
methods, and applications in interpretable machine learning’. Proceedings
of the National Academy of Sciences of the United States of America.
2019;116(44):22071–80.
[6] Pedregosa F., Varoquaux G., Gramfort A., et al. ‘Scikit-­learn: machine learn-
ing in python’. Journal of Machine Learning Research. 2011;12:2825–30.
[7] PyDotPlus pydotplus homepage [online] [PyDotPlus Developers]. 2022 Jul
15. Available from https://pydotplus.readthedocs.io/reference.html
[8] ​Sklearn.​tree.​deci​sion​tree​regr​essor [online]. scikit-­learn developers. 2022 Jul
19. Available from https://scikit-learn.org/stable/modules/generated/sklearn.​
tree.DecisionTreeRegressor.html
 Explanation of black-­box ML models in maritime transport  207

[9] Lundberg S.M., Lee S.I. ‘A unified approach to interpreting model predic-
tions’. Advances in Neural Information Processing Systems. 2017;30.
[10] Lundberg S.M., Erion G.G., Lee S.I. ‘Consistent individualized feature
Attribution for tree ensembles’. ArXiv:180203888. 2018.
[11] Shap API reference [online] [Scott Lundberg Revision]. 2022 Aug 19.
Available from https://shap-lrjball.readthedocs.io/en/latest/api.html
This page intentionally left blank
 Chapter 15
Linear optimization

We will discuss linear optimization in this chapter, as linear optimization is impor-

tant in itself and is the foundation of more advanced optimization techniques.

15.1 Basics

Example 15.1. Consider a service with the port rotation below:

‍Melbourne (M, 1) ! Sydney (S, 2) ! Brisbane (B, 3) ! Melbourne (M, 1)‍

Ships with a capacity of 1000 (TEUs) are deployed to provide a weekly frequency. The
container shipment demand is: Melbourne to Sydney ‍qMS = 800,‍ Melbourne to Brisbane
MB
‍q = 700.‍ The profit of transporting 1 TEU from Melbourne to Sydney is $200, and
from Melbourne to Brisbane is $300. Develop an optimization model to evaluate the maxi-
mum profit ($/week) the company can make. ‍‍
To formulate the above problem as a mathematical model, first, we need decision
variables, or variables for short. Decision variables are generally what we can control.
Some models have auxiliary decision variables simply for the purpose of ease of model-
ing. In this example, the demand ‍qMS = 800‍is beyond our control and hence is a param-
eter for the model (a parameter is a given/known constant). Here, we can control how
many containers to transport from Melbourne to Sydney, denoted by ‍yMS ,‍ and how many
from Melbourne to Brisbane, denoted by ‍yMB.‍
Second, we need to define the objective function. The objective function is an expres-
sion of the target associated with the decision variables. Here, we choose the profit as the
objective function:

‍ 200yMS + 300yMB‍

We seek to find values for ‍yMS ‍ and ‍yMB ‍ that could maximize the objective
function.
Third, we need to define the constraints because the decision variables cannot take all
real values. For example, ‍yMB‍cannot be 1 billion because of limited demand and limited
ship capacity; ‍yMB‍cannot be ‍4‍, either.
210  Machine learning and data analytics for maritime studies

Solution. Let ‍yMS ‍and ‍yMB‍be the decision variables representing the volumes of con-
tainers transported from Melbourne to Sydney and Melbourne to Brisbane, respectively.
The model is:
max 200yMS + 300yMB (maximize the total profit)
subject to
yMS + yMB  1000
(ship capacity constraint on the leg from M to S)
yMB  1000
(cannot carry more containers than the ship capacity on the leg from S to B)
yMs  800
(cannot carry more containers for the OD pair (M S) than the demand)
yMB  700
(demand constraint for the OD pair (M, B))
yMS  0
(cannot carry a negative number of containers for the OD pair (M, S))
yMB  0
‍ (nonnegativity constraint on yMB ). ‍
We can see from the above solution that, in an optimization model we need to
(1) define the decision variables; (2) define the objective function and explain it;
and (3) define the constraints and explain each constraint. We can use any symbol
to represent a decision variable; however, we often use symbols that are easy to
remember. If the meaning of the objective function or a constraint is straightfor-
ward, it is acceptable not to explain it. In the above solution, “max” means we
want to maximize the objective function; we use “min” if we want to minimize the
objective function (e.g., when the objective function represents cost). The model
for Example 1 has two variables and six constraints.
Example 15.2. Reconsider the service in Example 15.1. Ships with a capacity of 1500
(TEUs) are deployed to provide a weekly frequency. The container shipment demand
is: Melbourne to Sydney ‍qMS = 800‍, Melbourne to Brisbane ‍qMB = 700‍, and Sydney
to Brisbane ‍qSB = 900‍. The profit of transporting 1 TEU from Melbourne to Sydney
is $200, from Melbourne to Brisbane is $300, and from Sydney to Brisbane is $100.
Develop an optimization model to evaluate the maximum profit ($/week) the company
can make.
Solution. Let yMS, yMB, and ySB be the decision variables representing the vol-
umes of containers transported from M to S, M to B, and S to B, respectively.
The model is:

‍ max 200yMS + 300yMB + 100ySB ‍

subject to
 Linear optimization  211

yMS + yMB  1500

(ship capacity constraint on the leg from M to S)
yMB + ySB  1500
(cannot carry more containers than the ship capacity on the leg from S to B)
yMs  800
yMB  700
ySB  900
yMS  0
yMB  0
‍ y  0.
SB
‍
Example 15.3. Reconsider Example 15.1. Among the 800 TEUs to transport from
Melbourne to Sydney, 200 TEUs are from a major customer and must be transported
(otherwise the customer may be lost in the future). Develop an optimization model to
evaluate the maximum profit ($/week) the company can make.
Solution. The answer is the model in the solution to Example 1 plus the following
constraint: ‍yMS  200.‍
Example 15.4. Reconsider Example 15.1. Suppose that if the company rejects 1 TEU from
Melbourne to Sydney, it loses $30; and if it rejects 1 TEU from Melbourne to Brisbane,
it loses $20 (e.g., loss of good will). Formulate a linear optimization model to help the
company make the decision of how to transport containers.
Solution.The answer is the model in the solution to Example 15.1, except that the
objective function is changed to

Figure 15.1   CC1 service by OOCL

212  Machine learning and data analytics for maritime studies

‍max 200y + 300y  30(800  y )  20(700  y ).‍

MS MB MS MB

Example 15.5. Consider the CC1 service in Figure 15.1 with the port rotation
Shanghai (S, 1) ! Kwangyang (K, 2) ! Pusan (P, 3) ! Los Angeles (L, 4)
‍ ! Oakland (O, 5) ! Pusan (P, 6) ! Kwangyang (K, 7) ! Shanghai (S, 1)‍
Ships with a capacity of 8000 (TEUs) are deployed to provide a weekly frequency. The
container shipment demand is: Shanghai to Los Angeles ‍qSL = 4500,‍ Kwangyang
to Los Angeles ‍qKL = 1000,‍ Pusan to Los Angeles ‍qPL = 1500,‍ Oakland to Shanghai
OS SP
‍q = 3700,‍ and Shanghai to Pusan ‍q = 1900.‍ The profit of transporting 1 TEU from
Shanghai to Los Angeles is $1800, Kwangyang to Los Angeles is $1900, Pusan to Los
Angeles is $1600, Oakland to Shanghai is $900, and Shanghai to Pusan is $500. Develop
an optimization model to evaluate the maximum profit ($/week) the company can make.
Solution. Let ySL, yKL, yPL, yOS, and ySP be the decision variables representing the volumes
of containers transported from S to L, K to L, P to L, O to S, and S to P, respectively. The
model is:
max1800ySL + 1900yKL + 1600yPL + 900yOS + 500ySP
subject to
ySL + ySP  8000
(ship capacity constraint on the leg from S to K)
ySL + yKL + ySP  8000
(ship capacity constraint on the leg from K to P)
ySL + yKL + yPL  8000
(ship capacity constraint on the leg from P to L)
yOS  8000
(ship capacity constraint on the leg from O to P)
yOS  8000
(ship capacity constraint on the leg from P to K)
yOS  8000
(ship capacity constraint on the leg from K to S)
0  ySL  4500
0  yKL  1000
0  yPL  1500
0  yOS  3700
‍ 0  ySP  1900. ‍
 Linear optimization  213

15.2  C
 lassification of linear optimization models according to
solutions

The examples in section 15.1 are all linear optimization models. A linear optimization
model has a linear objective function to be maximized or minimized, and a set of linear
constraints. The following functions are linear:
5x + 6y  3.4z
‍ 100x1  0.001x2 + 1002. ‍
The following functions are nonlinear:
5x + 6y2  3.4z
100x1  0.001 sin x2 + 1002
‍ 3xy. ‍
The following constraints are linear:
5x + 6y  3.4z  23
100x1  0.001x2 + 1002  100
‍ 100x1  0.001x2 + 1002 = 100.‍
It should be noted that we do not usually consider “<” or “>” constraints in linear
optimization, e.g., ‍x1 + x2 < 5‍.
A feasible solution to a linear optimization model is a vector of values of the deci-
sion variables that satisfies all the constraints. For example, in the model for Example 1,
MS MB
‍(y , y ) = (0, 1)‍ is a feasible solution. The set of all feasible solutions is called the
feasible set. We can calculate the objective function value for each feasible solution. For
example, the objective function value of the solution ‍(yMS , yMB ) = (0, 1)‍is 300. An opti-
mal solution is the “best” feasible solution, which has the largest objective function value
for a maximization model, and the smallest objective function value for a minimization
model. The objective function value of an optimal solution is called the optimal objective
function value.
Example 15.6. Write down three distinct feasible solutions to the model in the solution to
Example 1, and calculate their objective function values.
Solution. There are many feasible solutions. For example, the objective function value
of the solution ‍(yMS , yMB ) = (0, 1)‍ is 300; the objective function value of the solution
MS MB MS MB
‍(y , y ) = (0, 0)‍is 0; the objective function value of the solution ‍(y , y ) = (10, 10)‍
MS MB
is 5000; the objective function value of the solution ‍(y , y ) = (100, 100)‍is 50,000.
A linear optimization model can be classified into three categories according to solu-
tions: infeasible, unbounded, and having an optimal solution. A linear optimization model
may have no feasible solution, i.e., the model is infeasible, e.g.:
‍max x + y ‍
214  Machine learning and data analytics for maritime studies

subject to
x+y1
x0
‍ y  0.‍
A linear optimization model is infeasible if and only if its feasible set is empty, i.e., no
solution satisfies all the constraints simultaneously. Whether a linear optimization model
is infeasible or not has nothing to do with its objective function. A linear optimization
model either has 0, or 1, or an infinite number of feasible solutions. A linear optimization
model is feasible if it has at least one feasible solution.
A linear optimization model may be unbounded, i.e., for a maximization model, the
objective function value can be infinitely large and for a minimization model, the objective
function value can be infinitely small, e.g.:
‍max x + y ‍
subject to
x+y1
x0
‍ y  0.‍ 
Note that if a linear optimization model is unbounded, its feasible set must be
unbounded. If the feasible set of a linear optimization model with decision vari-
ablesx‍x11, ,xx22, ,   , ,xxnn‍ is bounded, i.e., there exists a positive number ‍ M ‍such that any
feasible solution satisfies  ‍ M  x1  M, M  x2  M,   , M  xn  M,‍ then
the model will not be unbounded. If the feasible set is unbounded, the model may
be bounded (e.g., minimizing ‍x ‍ subject to ‍x  0‍) or unbounded (e.g., minimizing
‍ x ‍subject to ‍x  0‍). For most practical problems, we do not worry about whether

they unbounded, because in reality the absolute values of the decision variables
cannot be infinitely large.
If a linear optimization model is feasible and not unbounded, then it has an
optimal solution, e.g.:
‍ max x + y ‍
subject to
x+y1
x1
‍ y  1.‍
We often add the superscript “‍‍” to a decision variable to represent its value in an
optimal solution. For example, we often use ‍(x , y )‍to represent the optimal values
of the decision variables ‍(x, y)‍.
If a linear optimization model has an optimal solution, the optimal solution may
not be unique, e.g.:
‍ max x + y ‍
 Linear optimization  215

subject to
x+y1
x1
‍ y  1.‍
If a linear optimization model has more than one optimal solution, then it has an
infinite number of optimal solutions. In most cases, we are only interested in obtain-
ing one of them.
If an optimization model has an optimal solution, then there is no loss of gen-
erality to say “suppose that it is a minimization model with a positive objective
function.” For example, if we aim to maximize ‍3x  4y ,‍ then we can minimize
(3x
‍  4y) + z ‍ subject to ‍z = 1010‍ (i.e., we let the new decision variable ‍ z ‍equal a
very large positive number).
Example 15.7. Suppose that all linear optimization models in this question have
optimal solutions. (1) Consider a minimization linear optimization model with
several constraints, and one of them is ‍3x  4y  5.‍ If the constraint is changed
to ‍3x  4y  6,‍ how will the feasible set change? How will the optimal objective
function value change? (2) Consider a linear optimization model with several con-
straints and minimizing ‍3x  4y .‍ If the objective function is changed to ‍6x  8y ,‍ how
will the feasible set change? How will the optimal solution change? How will the
optimal objective function value change?
Solution. (1) The feasible set will be larger or not change. The optimal objective
function value will be smaller or not change. (2) The feasible set will not change.
The optimal solution will not change. The optimal objective function value will be
twice as large as before.
Example 15.8. Reconsider the model in the solution to Example 1. (1) If the ship capac-
ity increases, how will the feasible set change? How will the optimal objective function
value change? (2) If the demand from Melbourne to Brisbane decreases, how will the
feasible set change? How will the optimal objective function value change? (iii) If the
profit of transporting 1 TEU from Melbourne to Sydney is $400 instead of $200, and from
Melbourne to Brisbane is $600 instead of $300, how will the feasible set change? How
will the optimal solution change? How will the optimal objective function value change?
Solution. (1) The feasible set will be larger or not change. The optimal objective
function value will be larger or not change. (2) The feasible set will be smaller or
not change. The optimal objective function value will be smaller or not change. (3)
The feasible set will not change. The optimal solution will not change. The optimal
objective function value will be twice as large as before.

15.3  Equivalence between different formulations

It is helpful to understand that some forms of a linear optimization model are equiva-
lent. When we say transform Model A to Model B, we mean that the Model B is
equivalent to Model A in the sense that gives an optimal solution to Model B, we can
easily derive an optimal solution to Model A.
216  Machine learning and data analytics for maritime studies

Regarding linear optimization, first, as we have seen, a maximization objec-

tive can be transformed to a minimization objective, and vice versa. For example,
maximizing ‍5x1  6x2‍ is equivalent to minimizing (5x ‍ 1  6x2 ).‍ Second, “
‍ ‍” and
“‍ ‍” constraints can be transformed to each other. For example, ‍5x1  6x2  10‍ is
equivalent to (5x
‍ 1  6x2 )  10.‍ Third, “=” constraints can be transformed to “ ‍
‍” and “‍ ‍” constraints. For example, ‍5x1  6x2 = 10‍is equivalent to the combination
of ‍5x1  6x2  10‍and ‍5x1  6x2  10‍.
Example 15.9. Transform the following linear optimization model to one with a
minimization objective and “ ‍ ‍” constraints.
‍ max 5x + 6y ‍ 
subject to
3x  4y  34
45x  98y = 9
x1
‍ y  0.‍
Solution
‍ min 5x  6y ‍ 
subject to
3x + 4y  34
45x  98y  9
(45x  98y)  9
x  1
‍ y  0.‍
Example 15.10. Transform the model in Example 9 to one with a maximization
objective and “
‍ ‍” constraints.
Solution
‍ max 5x + 6y ‍
subject to
3x  4y  34
45x  98y  9
(45x  98y)  9
x1
‍ y  0.‍
Fourth, “‍ ‍” and “
‍ ‍” constraints can be transformed to a combination of “=” con-
straints and nonnegativity constraints on the decision variables (i.e., constraints
imposing that the decision variables are nonnegative). For example, ‍5x1  6x2  10‍
is equivalent to the combination of ‍y = 10  (5x1  6x2 )‍and ‍y  0.‍
 Linear optimization  217

Fifth, we can transform a model to one in which all decision variables are non-
negative. For example, if we have a constraint ‍x  0‍, we can let ‍u = x ‍and replace
‍x ‍by ‍ u‍in the model. If we have a constraint ‍x  3‍, we can let ‍u = x  3‍and replace
‍x ‍ by ‍u + 3‍ in the model. If we have a constraint ‍x  3,‍ we can let ‍u = 3  x ‍ and
replace ‍x ‍ by ‍3  u‍ in the model. If the model does not specify whether ‍x ‍ is non-
negative or nonpositive, we can define ‍u1  0, u2  0‍ and replace ‍x ‍ by ‍u1  u2‍ in
the model.
Example 15.11. Transform the model in Example 5.9 to one with only “=” con-
straints and nonnegative decision variables.
Solution. As ‍x  1‍ and ‍y  0‍, we let ‍s = x  1‍ and ‍t = y ‍. Hence, ‍x = s + 1, s  0‍
and ‍y = t, t  0‍. The model is

‍ max 5(s + 1) + 6(t) (i.e., 5s  6t + 5)‍

subject to

3(s + 1)  4(t)  34 (i.e., 3s + 4t  31)

45(s + 1)  98(t) = 9 (i.e., 45s + 98t = 36)
s0
‍ t  0.‍

Letting ‍r = (3s + 4t)  31,‍ the model is equivalent to

‍max 5s  6t (note that the constant 5 does not affect the model)‍
subject to

3s + 4t  r = 31
45s + 98t = 36
r0
s0
‍ t  0.‍

Sixth, we can transform the objective function to one that is equal to a decision
variable. For example, maximizing ‍5x1  6x2‍is equivalent to maximizing ‍u‍, subject
to ‍u = 5x1  6x2.‍ It is also equivalent to maximizing ‍u‍, subject to ‍u  5x1  6x2.‍
In plain words, this property means that the objective function can somehow be
considered as a constraint; it also means that if we can solve an optimization model
with complex constraints, we can also solve an optimization model with complex
constraints and a complex objective function.
Finally, in some problems there is no objective function, for instance, when
we are only interested in finding a feasible solution. In this case, we can aim to
“min 0.”
Two general forms of linear optimization models are frequently used in theoret-
ical analysis: the canonical form that maximizes ‍cT x ‍subject to ‍Ax  b‍and ‍x  0‍,
218  Machine learning and data analytics for maritime studies

and the standard form that maximizes ‍cT x ‍ subject to ‍Ax = b‍ and ‍x  0‍, where ‍x ‍
is a column vector representing ‍n‍ decision variables, ‍ A‍is an ‍m  n‍ matrix, ‍c ‍ is a
column vector with ‍n‍rows, ‍b‍is a column vector with ‍m‍rows, and ‍ A‍, ‍b‍, and ‍c ‍are
all parameters.
Example 15.12. Transform the model in Example 15.9 to the canonical form.
Solution. Letting ‍x = u  v, u  0, v  0‍, and ‍y = w, w  0,‍ the model is trans-
formed to

‍max 5(u  v) + 6(w)‍

subject to

3(u  v) + 4(w)  34

45(u  v)  98(w)  9
45(u  v) + 98(w)  9
(u  v)  1
u0
v0
‍ w  0.‍
2 3
3 3 4
6 7
6 45 45 98 7
Hence, in the canonical form, x = [ u v w ]T , A =6 7
‍ ‍ 6 45 45 98 7,
4 5
‍ 1 1 0 ‍

b = [ 34 9 9 1 ]T , and c = [ 5 5 6 ]T , with ‍x = u  v ‍ and

‍ ‍ ‍ ‍
‍y = w .‍

15.4  Graphical method for models with two variables

Linear optimization models with two variables can be solved intuitively using
graphs. In practical applications, hardly any problem has only two variables.
However, learning the graphical method is helpful for appreciating the properties of
linear optimization models.
In the graphical method, we first draw constraints that define an upper or lower
bound for a decision variable (e.g., ‍x  0, y  100‍), then draw the other constraints,
and finally draw a line or a series of parallel lines that represent the objective
function.
Example 15.13. Consider the model below:
‍ max x1 + x2‍
 Linear optimization  219

Figure 15.2   Graphical method for Example 15.13

subject to
x1  x2
x1  1
x2  1
x1  0
‍ x2  0.‍
Use the graphical method to find the optimal solution.
Solution. See Figure 15.2. The optimal solution corresponds to the intersection of
the lines ‍x1 = x2‍ and ‍x1 = 1‍. Therefore, the optimal solution is ‍x1 = 1, x2 = 1‍. The
optimal objective function value is 2.
Example 15.14. Consider the model below:
‍ min x2  x1‍
subject to
x1  x2
x1  1
x2  1
x1  0
‍ x2  0.‍ 
Use the graphical method to find the optimal solution.
220  Machine learning and data analytics for maritime studies

Figure 15.3   Graphical method for Example 15.14

Solution. See Figure 15.3. The optimal solution corresponds to the intersection

of the lines ‍x1 = 1‍and ‍x2 = 0‍. Therefore, the optimal solution is ‍x1 = 1, x2 = 0.‍ The
optimal objective function value is ‍1‍.
Example 15.15. Consider the model below:
‍ max 2x1 + x2‍
subject to
x1 + x2  1
x1  2
x2  1
x1  0
‍ x2  0.‍ 
Use the graphical method to find the optimal solution.
Solution. See Figure 15.4. The optimal solution corresponds to the intersection of
the lines ‍x1 = 2‍ and ‍x2 = 1‍. Therefore, the optimal solution is ‍x1 = 2, x2 = 1‍. The
optimal objective function value is ‍5‍.
Example 15.16 Consider the model below:
‍ max 2x1 + x2‍
 Linear optimization  221

Figure 15.4   Graphical method for Example 15.15

subject to
x1 + x2  1
x1  3
x2  1
x1  0
‍ x2  0.‍ 
Use the graphical method to find the optimal solution.
Solution. See Figure 15.5. The optimal solution corresponds to the intersection of
the lines ‍x1 = 3‍ and ‍x2 = 1‍. Therefore, the optimal solution is ‍x1 = 3, x2 = 1.‍ The
optimal objective function value is ‍7‍.
Example 15.17. Consider the model below:
‍ max 2x1 + x2‍
subject to
x1  1
x2  1
x1 + x2 = 1
x1  0
‍ x2  0.‍
Use the graphical method to find the optimal solution.
222  Machine learning and data analytics for maritime studies

Figure 15.5   Graphical method for Example 15.16

Solution. See Figure 15.6. Note that the feasible set is a line segment. The optimal
solution corresponds to the intersection of the lines ‍x1 + x2 = 1‍and ‍x1 = 1‍. Therefore,
the optimal solution is ‍x1 = 1, x2 = 0.‍ The optimal objective function value is 2.
Example 15.18. Consider the model below:

Figure 15.6   Graphical method for Example 15.17

 Linear optimization  223

Figure 15.7   Graphical method for Example 15.18

‍ max x1‍

subject to

2x1 + x2  1000
3x1 + 4x2  2400
x1 + x2  700
x1  x2  350
x1  0
‍ x2  0.‍

Use the graphical method to find the optimal solution.

Solution. See Figure 15.7. The optimal solution corresponds to the intersection
of the lines ‍2x1 + x2 = 1000‍ and ‍x1  x2 = 350‍. Therefore, the optimal solution is
 
‍x1 = 450, x2 = 100.‍ The optimal objective function value is 450.
224  Machine learning and data analytics for maritime studies

Figure 15.8   Graphical method for Example 15.19

Example 15.19. Consider the model below:

‍ max 8x1 + 5x2‍

subject to

2x1 + x2  1000
3x1 + 4x2  2400
x1 + x2  700
x1  x2 = 350
x1  0
‍ x2  0.‍

Use the graphical method to find the optimal solution.

Solution. See Figure 15.8. Note that the feasible set is a line segment. The opti-
mal solution corresponds to the intersection of the lines ‍2x1 + x2 = 1000‍ and
 
‍x1  x2 = 350‍. Therefore, the optimal solution is ‍x1 = 450, x2 = 100‍. The optimal
objective function value is 4100.
 Linear optimization  225

We can see from the above examples that if a linear optimization model has an
optimal solution, then there exists an optimal solution that is at a “corner point” of
the feasible set*.
The graphical method can also be used to identify whether a linear opti-
mization model is infeasible, unbounded, or has an infinite number of optimal
solutions.

15.5  Using software to solve linear optimization models

Linear optimization models can be solved by simplex method, ellipsoid method, or

interior point method.
When we say “solve a linear optimization model,” we mean finding an optimal
solution to it, or proving that it is infeasible or unbounded. If we know an optimal
solution, it is very easy to calculate the optimal objective function value. Knowing
the optimal objective function value without knowing an optimal solution is of little
value in most cases.
There are many free and commercial solvers for linear optimization models, e.g.,
Excel, MATLAB, lpSolve, GUROBI, and CPLEX. A linear optimization solver is
a software package that can do the following things for a linear optimization model:
(1) check whether it is infeasible; (2) check whether it is unbounded; and (3) return
an optimal solution if there is one. Note that if a linear optimization model has more
than one optimal solution, a solver generally returns only one optimal solution that
is at a corner point.
Generally speaking, linear optimization models are “easy” to solve: models
with 100,000 variables and 100,000 constraints can be solved in 1 minute†. Since
free solvers can also solve linear optimization models very efficiently, hardly any
commercial solver is solely dedicated to linear optimization models. In fact, com-
mercial solvers are mainly aimed to solve the much more difficult mixed-­integer
linear optimization models.
We can learn how to use Excel to solve a few simple linear optimization models
to appreciate the use of solvers. Since the solvers are updated frequently, we do not
discuss the detailed steps. Readers can easily Google “Excel linear optimization” to
learn more about how to configure Excel to solve linear optimization models. Note
that to use Excel, we need the “Solver Add-­in” function.
Example 15.20. Use a solver (e.g., Excel) to solve the following linear optimization
model:
‍ max 200x + 300y ‍

* 
There are some linear optimization models with no corner points, e.g., maximizing 0x + 2y subject to0
≤ y ≤ 1. We generally do not need to worry about them in practical applications.
† 
The exact time depends on the parameters of the model, the solver, and the computer.
226  Machine learning and data analytics for maritime studies

subject to
x + y  1000
y  1000
x  800
y  700
x0
‍ y  0. ‍
Solution. The optimal solution is ‍x = 300, y = 700‍.
Example 15.21. Use a solver (e.g., Excel) to solve the following linear optimization
model:
‍ max 200x + 300y ‍
subject to
x + y  1000
y  1000
x  800
y  700
x0
‍ y  0. ‍
Solution.If Excel is used, then the results are “The Objective Cell values do not con-
verge. Solver can make the Objective Cell as large (or as small when minimizing) as
it wants,” which indicates that the model is unbounded.
Example 15.22. Use a solver (e.g., Excel) to solve the following linear optimization
model:
‍max 200x + 300y ‍
subject to
x + y  1000
y  1000
x  800
y  700
x0
‍ y  0. ‍
Solution. If Excel is used, the results are “Solver could not find a feasible solution.
Solver cannot find a point for which all Constraints are satisfied,” which indicates
that the model is infeasible.
Example 15.23. Use a solver (e.g., Excel) to solve the following linear optimization
model:

‍ max x + y ‍
 Linear optimization  227

subject to
x+y1
x0
‍ y  0.‍ 

Solution. Although there are an infinite number of optimal solutions, Excel only
provides one optimal solution, e.g., ‍x = 1, y = 0.‍

15.6  An in-depth understanding of linear optimization

The following questions are on linear optimization models. For each question, you
should either answer that such a linear optimization model does not exist (and you
do not need to provide the reason), or give an example of such a linear optimization
model.
Example 15.24. Give a model that is both infeasible and unbounded.
Solution. Such a model does not exist.
Example 15.25. Give a model whose optimal objective function value is 1, and after
removing one constraint, the optimal objective function value is 0.
Solution. Minimizing ‍x‍subject to ‍x  1, x  0.‍ After removing the constraint ‍x  1,‍
the optimal objective function value is 0.
Example 15.26.
1. A model is infeasible, and after removing one constraint, it is feasible.
2. A model is infeasible, and after removing one constraint, it is unbounded.
3. A model is infeasible, and after removing one constraint, it has an optimal
solution.
4. A model is infeasible, and after removing one constraint, it has an infinite num-
ber of optimal solutions.
5. A model has an optimal solution, and after adding one constraint, it is
unbounded.
6. A model has an optimal solution, and after adding one constraint, it is
infeasible.
7. A model has exactly one optimal solution, and after adding one constraint, it
has an infinite number of optimal solutions.
8. A model has an infinite number of optimal solutions, and after adding one con-
straint, it has exactly one optimal solution.
9. A model is infeasible, and after changing its objective function, it has an opti-
mal solution.
10. A model has an optimal solution, and after changing its objective function, it is
infeasible.
11. A model has an optimal solution, and after changing its objective function, it is
unbounded.
12. A model is unbounded, and after changing its objective function, it has exactly
one optimal solution.
228  Machine learning and data analytics for maritime studies

13. A model is unbounded, and after changing its objective function, it has an infi-
nite number of optimal solutions.

Solution.

1. Maximize ‍x ‍subject to ‍x  1‍and ‍x  0‍. Remove the first constraint.

2. Minimize ‍x ‍subject to ‍x  1‍and ‍x  0‍. Remove the first constraint.
3. Maximize ‍x ‍subject to ‍x  1‍and ‍x  0‍. Remove the first constraint.
4. Maximize ‍0‍subject to ‍x  1‍and ‍x  0‍. Remove the first constraint.
5. Not possible.
6. Minimize ‍x ‍subject to ‍x  1‍. Add ‍x  0‍.
7. Maximize ‍x ‍subject to ‍x  0‍, ‍y  0‍, and ‍x + y  1‍. Add ‍x  0.5‍.
8. Maximize ‍x ‍subject to ‍x  0‍, ‍y  0‍, ‍x + y  1‍, and ‍x  0.5‍. Add ‍x + y  0.5‍.
9. Not possible.
10. Not possible.
11. Minimize ‍x ‍subject to ‍x  0‍. Change the objective function to ‍max x ‍.
12. Maximize ‍x ‍subject to ‍x  0‍. Change the objective function to ‍min x ‍.
13. Maximize ‍x + y ‍ subject to ‍x  0‍, ‍y  0‍, and ‍x + y  1‍. Change the objective
function to ‍min x + y ‍.

15.7  Useful applications of linear optimization solvers

Although a linear optimization solver only provides one optimal solution if there
are an infinite number of optimal solutions, we can take advantage of the solver in a
number of smart ways to address more problems.
Example 15.27. Given a linear optimization model with a non-­empty and bounded
feasible set defined by ‍Ax  b‍and ‍x  0‍, where the vector of decision variables ‍x ‍
has ‍n‍ elements and ‍ A‍, ‍b‍, and ‍c ‍ are parameters of appropriate dimensions, how to
use a linear optimization solver to check whether it has only one feasible solution or
an infinite number of feasible solutions.
Solution. The number of feasible solutions is independent of the objective function.
We solve the following two models: maximizing x1 subject to ‍Ax  b‍ and ‍x  0‍,
and minimizing x1 subject to ‍Ax  b‍ and ‍x  0‍. Here both models have an opti-
mal solution. If their optimal objective function values are different, then there are
an infinite number of feasible solutions to the original model. Otherwise we check
models that maximize and minimize x‍x22,,xx33,, ,,xxnn‍. If all of the optimal objective
function values of the ‍n‍ maximization models are the same as the corresponding
values of the minimization models, there is only one feasible solution.
Example 15.28. Given a linear optimization model that is feasible and not
unbounded: maximizing ‍cT x ‍subject to ‍Ax  b‍and ‍x  0‍, where the vector of deci-
sion variables ‍x ‍ has ‍n‍ elements and ‍ A‍, ‍b‍ and ‍c ‍ are parameters of appropriate
dimensions, how to use a linear optimization solver to check whether it has only one
optimal solution or an infinite number of optimal solutions.
 Linear optimization  229

Solution. Obtain an optimal solution, denoted by ‍x‍. Now the problem becomes how
to check whether there is one or an infinite number of feasible solutions to the set of
constraints ‍Ax  b,‍ ‍x  0,‍ and ‍cT x = cT x‍, which is Example 15.27.
Example 15.29. Given a set of constraints ‍Ax  b‍ and ‍x  0,‍ where the vector of
decision variables x = [ x y ]T and ‍ A‍, ‍b‍, and ‍c ‍ are parameters of appropriate
‍ ‍
dimensions. We want to find the optimal values of the decision variables satisfy-
ing two conditions. Condition 1: the expression ‍3x  4y ‍is maximized. Condition 2:
among all the feasible solutions satisfying Condition 1, we want to find the one that
minimizes ‍2x + 3y .‍ How to use a linear optimization solver to address this problem‡
Solution. The first approach is to use the big-­M method. Our first priority is to maxi-
mize ‍3x  4y ,‍ and our second priority is to maximize 2x ‍  3y .‍ Therefore, we can
solve the model maximizing 10 8 (3x  4y) + (2x  3y) subject to Ax  b and x  0.
‍ ‍ ‍ ‍ ‍ ‍
Of course, the solution may not be accurate.
The second approach involves sequential optimization. We first solve the model
maximizing ‍3x  4y ‍ subject to ‍Ax  b‍ and ‍x  0‍ to obtain the optimal solution
‍(Ox, yO ).‍ We then solve the model minimizing ‍2x + 3y ‍ subject to ‍Ax  b,‍ ‍x  0,‍ and
 
‍3x  4y = 3Ox  4Oy .‍ Its optimal solution ‍(x , y )‍is what we want.
Example 15.30. A linear optimization model has decision variablesx‍x1 ,1 ,xx2 ,2 ,   , ,xxn‍nand
the following constraints:
a11 x1 + a12 x2 + ... + a1n xn  b1
a21 x1 + a22 x2 + ... + a2n xn  b2
..
.
am1 x1 + am2 x2 + ... + amn xn  bm
0  x1  1
0  x2  1
..
.
‍ 0  xn  1 ‍
where ‍aij , bi ‍ are all constants, i‍ = 1, 2,   , m
, m,‍ ‍j = 1, 2,   , n
, n.‍ We do not know its
objective function yet. How to use a linear optimization solver to check whether
the first constraint ‍a11 x1 + a12 x2 +    + a1n xn  b1‍ is redundant (i.e., removing the
constraint does not affect the model)?
Solution. We actually need to check whether the feasible set defined by all the con-
straints is the same as the feasible set defined by the constraints excluding the first
one. To this end, we solve the model minimizing ‍a11 x1 + a12 x2 +    + a1n xn‍subject
to the second to the last constraints. If the optimal objective function value is not
smaller than b1, which means, any point that satisfies the second to the last con-
straints automatically satisfies the first one, then the first constraint is redundant;
otherwise it is not redundant.

For example, in traffic control, Condition 1 might mean maximizing the survival rates in road acci-
‡ 

dents, and Condition 2 might mean minimizing the travel delay due to road accidents. The objective in
Condition 1 is much more important than that in Condition 2.
12 Superconducting magnetic energy storage in power grids
230  Machine learning and data analytics for maritime studies
different cases. These cases contemplate unbalance in voltage, harmonic distortion
in voltage,15.31.
Example and power fluctuations
Suppose that Ω1 iscaused the high(‍xxvariation
byvectors
a set of 11, ,xx22, ,    , ,xxnn‍)weather
of defined resources.
by the fol-
lowing m + 2n constraints:
a11 x1 + a12 x2 + ... + a1n xn  b1
3.3 Conclusion
a21 x1 + a22 x2 + ... + a2n xn  b2
..
This chapter deals with some basics . of SMES and its control methodology. SMES is
one ofathe most developing and efficient energy storage devices. The integration of
m1 x1 + am2 x2 + ... + amn xn  bm
SMES systems in the AC power microgrids under connected operation mode allows
0  x1  1
compensating active and reactive power dynamically, which clearly improves the
grid performance in terms 0 ofx2power  1 factor reducing at the same time the power
oscillations produced by renewable .. generators. Different topologies of the VSC and
.
CSC systems are explored. The different control methodologies for VSC and CSC
are 0  xn  in
‍ used to mitigate the variation 1 voltage
‍ and power for grid-connected systems
‍ 2‍is aSMES.
using
 set of vectors(x ‍(x11,,xx22,,    ,,xxnn))‍defined by the following ‍m + 2n‍constraints:
c11 x1 + c12 x2 +    + c1n xn  d1
c21 x1 + c22 x2 +    + c2n xn  d2
..
.
cm1 x1 + cm2 x2 +    + cmn xn  dm
0  x1  1
0  x2  1
..
.
‍ 0  xn  1 ‍ 
where ‍aij , bi , cij , di ‍are all constants, i‍ = 1, 2,    , m
, m,‍ ‍j = 1, 2,    , n
, n.‍ How to use a lin-
ear optimization solver to check whether the two sets  ‍ 1‍ and ‍ 2‍are the same?
Solution. We first need to check whether the two sets are empty (minimizing 0 sub-
ject to the constraints that define each set). If  ‍ 1 = ¿‍and  ‍ 2 = ¿,‍ they are the same.
If exactly one of them is empty, they are different.
References
If ‍ 1 ¤ ¿‍ and  ‍ 2 ¤ ¿,‍ we refer to Example 15.30: if each constraint that
defines  ‍ 1‍ is redundant for 
‍ 2‍(i.e., any point in  automatically satisfies all con-
‍ 2‍for
[1] Farahani M. ‘A new control strategy of SMES mitigating subsynchronous
straints in  ‍ 1‍ ) , and each constraint that defines is
‍ 2‍34–39.
 redundant for  ‍ 1,‍ then the two
oscillations’. Physica C. 2012, vol. 483, pp.
sets are the same; otherwise they are different.
[2] Gil-González W., Montoya O.D., Garces A. ‘Control of a SMES for miti-
gating subsynchronous oscillations in power systems: a PBC-PI approach’.
Journal of Energy Storage. 2018, vol. 20, pp. 163–72.
[3] Ortega A., Milano F. ‘Generalized model of VSC-based energy storage sys-
tems for transient stability analysis’. IEEE Transactions on Power Systems.
2018, vol. 31(5), pp. 3369–80.
[4] Ali M.H., Murata T., Tamura J. ‘Transient stability enhancement by fuzzy
logic-controlled SMES considering coordination with optimal reclosing of
circuit breakers’. IEEE Transactions on Power Systems. 2018, vol. 23(2), pp.
631–40.
 Chapter 16
Advanced linear optimization

Basic linear optimization models have been introduced in Chapter 15. In this chap-
ter, more advanced linear optimization models will be covered.

16.1  Network flow optimization

Many practical problems are associated with the optimization of flow in a network,
such as transportation, telecommunication, and power transmission networks. A net-
work has nodes and arcs (arcs can also be called links). For example, a city logistics
network has intersections (nodes) and roads (arcs). Arcs are directional, i.e., cargoes/
passengers can only be transported from the tail node to the head node of an arc.
Therefore, a two-­way street is usually considered as two arcs.
Example 16.1: Walmart has three warehouses (W1, W2, and W3) that store
the same type of product and five supermarkets (S1–S5) that need the products
in a city. The number of products available at each warehouse, the number of
products needed at each supermarket, and the transportation cost per unit prod-
uct from each warehouse to each supermarket are shown below. Develop a linear
optimization model to help Walmart make the decision of how to transport the
products.

Warehouse Number of products available

W1 100
W2 200
W3 50

Supermarket Number of products needed

S1 80
S2 90
S3 70
S4 60
S5 50
232  Machine learning and data analytics for maritime studies

Unit cost ($) S1 S2 S3 S4 S5

W1 1 2 4 3 6
W2 5 2 4 4 4
W3 5 1 1 3 2

Solution. Let ‍fij ‍ be the decision variables representing the number of products
transported from warehouse i‍ = 1, 2, 3‍ to supermarket ‍j = 1, 2, 3, 4, 5.‍ The model is
as follows:
min f11 + 2f12 + 4f13 + 3f14 + 6f15 + 5f21 + 2f22 +
‍ 4f23 + 4f24 + 4f25 + 5f31 + f32 + f33 + 3f34 + 2f35 ‍ 
subject to
f11 + f12 + f13 + f14 + f15  100
f21 + f22 + f23 + f24 + f25  200
f31 + f32 + f33 + f34 + f35  50
f11 + f21 + f31 = 80
f12 + f22 + f32 = 90
f13 + f23 + f33 = 70
f14 + f24 + f34 = 60
f15 + f25 + f35 = 50
‍ fij  0, i = 1, 2, 3, j = 1, 2, 3, 4, 5.‍ 
Sometimes, we may use symbols to simplify the notation. We can define sets
‍I = f1, 2, 3g, J = f1, 2, 3, 4, 5g,‍ represent by pi the number of products available in
warehouse i‍ 2 I ‍, denote by qj the number of products needed by supermarket ‍j 2 J ,‍
and let ‍cij ‍be the cost of transporting one product from warehouse i‍ 2 I ‍to supermar-
ket ‍j 2 J .‍ Note that pi, qj, and ‍cij ‍ are all known parameters. Let ‍fij ‍ be the decision
variables representing the number of products transported from warehouse i‍ 2 I ‍ to
supermarket ‍j 2 J .‍ The model is as follows:
PP
min cij fij
‍ i2I j2J ‍
subject to
P
fij  pi , i 2 I
j2J
P
fij = qj , j 2 J
i2I

‍ fij  0, i 2 I, j 2 J.‍
It can be seen that the above model is very compact. Moreover, it is very general: we
might save the two sets ‍ I ‍ and ‍J ‍ and parameters pi, qj, and ‍cij ‍ in files, and program
the model using a linear optimization solver. Next time when Walmart plans the
transportation, it only needs to change the input files. The computer codes for linear
 Advanced linear optimization  233

optimization, which might be too complex for logistics managers to understand, do

not need to be changed.
Example 16.2: A small furniture company has three stores (S1–S3) and two
warehouses (W1 and W2). Stores 1, 2, and 3 need 12, 13, and 14 desks, respectively.
Warehouses 1 and 2 have 20 and 25 desks, respectively. The transportation costs
per desk from each warehouse to each store are shown below. Assuming that the
number of desks can be a fraction, develop a linear optimization model to help the
company make the decision of how to transport the desks.

Unit cost ($) S1 S2 S3

W1 1 2 4
W2 5 2 4

Solution. Let ‍fij ‍ be the decision variables representing the number of desks
transported from warehouse i‍ = 1, 2‍to store ‍j = 1, 2, 3‍. The model is as follows:
‍ min f11 + 2f12 + 4f13 + 5f21 + 2f22 + 4f23‍ 
subject to
f11 + f12 + f13  20
f21 + f22 + f23  25
f11 + f21 = 12
f12 + f22 = 13
f13 + f23 = 14
‍ fij  0, i = 1, 2, j = 1, 2, 3.‍ 
Example 16.3: This is a maximum flow problem. We have a crude oil pipeline
network shown in Figure  16.1. The arrows are the pipelines: crude oil can only
flow in the direction of the arrows, and the numbers on the arrows are the capacities
(maximum flow rates of crude oil) of the pipelines (1 000 tons/h). Node A is an oil
field and node E is a refinery factory. Develop a linear optimization model to find the
maximum flow of crude oil per hour from A to E.
In network flow problems, we can use the amount of cargo flow on each arc as
the decision variables. Such a formulation often needs the flow conservation equa-
tions: if a node does not export or import cargoes, then the total cargo inflow must

Figure 16.1   A crude oil pipeline network

234  Machine learning and data analytics for maritime studies

be equal to the total outflow; if the node exports cargoes, then the difference between
the total outflow and the total inflow equals the number of exported cargoes; and if
the node imports cargoes, then the difference between the total inflow and the total
outflow equals the number of imported cargoes. We first present a link flow formula-
tion for the maximum flow problem:
Solution. Let ‍N := fA, B, C, D, Eg‍be the set of nodes and ‍A ‍be the set of arcs:
‍ A := f(A, B), (A, D), (B, C), (B, D), (C, E), (D, C), (D, E)g.‍
Let ‍fij ‍be the decision variables representing the flow on arcs ‍(i, j) 2 A ‍. The model
is as follows:
‍ max fAB + fAD (maximize the total net outflow from node A)‍ 
subject to
fAB = fBC + fBD
(flow conservation at node B)
fBC + fDC = fCE
(flow conservation at node C)
fAD + fBD = fDC + fDE
(flow conservation at node D)
fAB  12, fAD  4, fBC  6, fBD  2
fCE  5, fDC  23, fDE  7
‍ fij  0, 8(i, j) 2 A. ‍
Note that in the above link flow formulation, the flow conservation equations at
nodes B, C, and D ensure that the total inflow to node E, ‍fCE + fDE ‍, equals the total
outflow from node A, ‍fAB + fAD ‍.
Example 16.4: In Example 16.3, develop a link flow linear optimization model
to find the maximum flow of crude oil per hour from node B to node E.
Solution. Let ‍N := fA, B, C, D, Eg‍be the set of nodes and ‍A ‍be the set of arcs:
‍ A := f(A, B), (A, D), (B, C), (B, D), (C, E), (D, C), (D, E)g.‍
Let fij be the decision variables representing the flow on arcs ‍(i, j) 2 A .‍ The model
is as follows:
‍ max fBC + fBD  fAB (maximize the total net outflow from node B)‍
subject to
fAB + fAD = 0 (flow conservation at node A)
fBC + fDC = fCE (flow conservation at node C)
fAD + fBD = fDC + fDE (flow conservation at node D)
fAB  12, fAD  4, fBC  6, fBD  2
fCE  5, fDC  23, fDE  7
‍ fij  0, 8(i, j) 2 A. ‍
 Advanced linear optimization  235

Note that as we seek the maximum flow from B to E, it does not make sense to have
flow to B. Therefore, it is correct to say that ‍fAB = 0‍and hence ‍fAB‍can be removed
from the model.
We can also use path flows as the decision variables in the maximum flow prob-
lem. To this end, we first need to enumerate all paths from the origin node to the
destination node.
Example 16.5: In Example 16.3, develop a path flow linear optimization model
to find the maximum flow of crude oil per hour from node A to node E.
Solution. From node A to node E, we have the following paths:
Path 1: A ! B ! C ! E
Path 2: A ! B ! D ! E
Path 3: A ! B ! D ! C ! E
Path 4: A ! D ! E
‍ Path 5: A ! D ! C ! E. ‍
Let fi be the decision variables representing the flow on path i ∈ {1, 2, 3, 4, 5}. The
path flow formulation is as follows:
X
5

max fi (maximize the total flow from node A to node E )

‍ i=1 ‍
subject to
f1 + f2 + f3  12 (capacity on arc (A, B))
f4 + f5  4 (capacity on arc (A, D))
f1  6 (capacity on arc (B, C))
f2 + f3  2 (capacity on arc (B, D))
f1 + f3 + f5  5 (capacity on arc (C, E))
f3 + f5  23 (capacity on arc (D , C))
f2 + f4  7 (capacity on arc (D, E))
‍ fi  0, i = 1, 2, 3, 4, 5. ‍
It should be noted that in many real problems, the sizes of the networks are very
large, i.e., the networks contain a large number of nodes and links. Enumerating all
of the paths from one node to another is impossible.
Example 16.6: In Example 16.3, develop a path flow linear optimization model
to find the maximum flow of crude oil per hour from node B to node E.
Solution. From node B to node E, we have the following paths:
Path 1 : B ! C ! E
Path 2 : B ! D ! E
‍ Path 3 : B ! D ! C ! E.‍ 
236  Machine learning and data analytics for maritime studies

Let fi be the decision variable representing the flow on path. The model is as follows:
X
3

max fi
‍ i=1 ‍
subject to
f1  6 (capacity on arc (B, C))
f2 + f3  2 (capacity on arc (B, D))
f1 + f3  5 (capacity on arc (C, E))
f3  23 (capacity on arc (D, C))
f2  7 (capacity on arc (D, E) )
‍ fi  0, i = 1, 2, 3. ‍
Example 16.7: Similar to Example 16.3, develop a link flow linear optimization
model to find the maximum flow of crude oil per hour from node A to node E in the
network shown in Figure 16.2.
Solution. Let ‍N := fA, B, C, D, Eg‍be the set of nodes and ‍A ‍be the set of arcs:

‍ A := f(A, B), (A, D), (B, C), (B, D), (C, A), (C, E), (D, C), (D, E), (E, C)g‍.
Let fij be the decision variables representing the flow on arcs ‍(i, j) 2 A .‍ The model
is as follows:
‍ max fAB + fAD  fCA‍ 
subject to
fAB = fBC + fBD (flow conservation at node B)
fBC + fDC + fEC = fCA + fCE (flow conservation at node C)
fAD + fBD = fDC + fDE (flow conservation at node D)
fAB  12, fAD  4, fBC  6, fBD  2
fCA  1, fCE  5, fDC  23, fDE  7, fEC  3
‍ fij  0, 8(i, j) 2 A. ‍

Figure 16.2   A crude oil pipeline network

 Advanced linear optimization  237

Example 16.8: This is a minimum cost flow problem. Consider a multimodal

transportation problem. Multimodal transportation means transportation with
more than one transportation mode, e.g., a combination of trains and trucks.
Figure  16.3 shows a multimodal transportation network including rail services
and truck services. There are four roads for trucks and three railway lines. The
maximum number of cargoes that can be transported on a railway line and on a
road is 10 and 5 units, respectively. The cargoes can only be transported in the
directions of the arrows. The numbers on the arrows represent the length of the
railway lines and roads. The transportation cost per unit distance is 3 for trucks
and 2 for trains. Node A represents the port of Los Angeles and node E is Chicago.
Develop a linear optimization model to find the minimum cost for transporting 12
units of cargoes from A to E.
We can see that both the maximum flow problem and the minimum cost flow
problem involve the determination of flow on each arc. Hence, their decision vari-
ables can be the same.
Solution. Let ‍N := fA, B, C, D, Eg‍be the set of nodes and ‍A ‍be the set of arcs:
‍ A := f(A, B), (A, D), (B, C), (B, D), (C, E), (D, C), (D, E)g.‍ 
Let ‍fij ‍be the decision variables representing the flow on arcs ‍(i, j) 2 A ‍. The model
is as follows:
‍ min 2  (6fAB + 5fCE + 8fDC ) + 3  (4fAD + 6fBC + 2fBD + 7fDE )‍
subject to
fAB + fAD = 12 (flow conservation at node A)
fAB = fBC + fBD (flow conservation at node B)
fBC + fDC = fCE (flow conservation at node C)
fAD + fBD = fDC + fDE (flow conservation at node D)
‍ fCE + fDE = 12 (flow conservation at node E)
fAB  10, fCE  10, fDC  10
fAD  5, fBC  5, fBD  5, fDE  5
fij  0, 8(i, j) 2 A.

Figure 16.3   A multimodal transportation network

 fAB + fAD = 12 (flow conservation at node A)
fAB = fBC + fBD (flow conservation at node B)
fBC + fDC = fCE (flow conservation at node C)
fAD + fBD = fDC + fDE (flow conservation at node D)
238  Machine learning and data analytics for maritime studies
fCE + fDE = 12 (flow conservation at node E)
fAB  10, fCE  10, fDC  10
fAD  5, fBC  5, fBD  5, fDE  5
fij  0, 8(i, j) 2 A. ‍
Example 16.9: In Example 16.8, develop a linear optimization model to find the
minimum cost for transporting 12 units of cargoes from node B to node E.
Solution. Let ‍N := fA, B, C, D, Eg‍be the set of nodes and ‍A ‍be the set of arcs:

‍ A := f(A, B), (A, D), (B, C), (B, D), (C, E), (D, C), (D, E)g.‍ 
Let ‍fij ‍be the decision variables representing the flow on arcs ‍(i, j) 2 A ‍. The model
is as follows:
‍ min 2  (6fAB + 5fCE + 8fDC ) + 3  (4fAD + 6fBC + 2fBD + 7fDE )‍ 
subject to
fAB + fAD = 0 (flow conservation at node A)
fBC + fBD  fAB = 12(flow conservation at node B)
fBC + fDC = fCE (flow conservation at node C)
fAD + fBD = fDC + fDE (flow conservation at node D)
fCE + fDE = 12(flow conservation at node E)
fAB  10, fCE  10, fDC  10,
fAD  5, fBC  5, fBD  5, fDE  5
‍ fij  0, 8(i, j) 2 A. ‍
Example 16.10: In Example 16.8, the price of a unit of cargo in Los Angeles is
100, and 115 in Chicago (the price has the same unit as the transportation cost). A
company buys cargoes in Los Angeles, transports them to Chicago, and sells them
to make a profit. Develop a linear optimization model to find the optimal number of
cargoes to purchase in Los Angeles that maximizes the profit.
Solution. Let ‍N := fA, B, C, D, Eg‍be the set of nodes and ‍A ‍be the set of arcs:

‍ A := f(A, B), (A, D), (B, C), (B, D), (C, E), (D, C), (D, E)g.‍ 
Let ‍fij ‍be the decision variables representing the flow on arcs ‍(i, j) 2 A .‍ Denote by ‍x ‍
the units of cargoes to purchase in Los Angeles. The model is as follows:
 
‍max(115  100)x  2  (6fAB + 5fCE + 8fDC ) + 3  (4fAD + 6fBC + 2fBD + 7fDE ) ‍ 

subject to

fAB + fAD = x (flow conservation at node A)

fAB = fBC + fBD (flow conservation at node B)
fBC + fDC = fCE (flow conservation at node C)
‍ fAD + fBD = fDC + fDE (flow conservation at node D)
fCE + fDE = x (flow conservation at node E)
fAB  10, fCE  10, fDC  10
fAD  5, fBC  5, fBD  5, fDE  5
x 0
fij  0, 8(i, j) 2 A.
 fAB + fAD = x (flow conservation at node A)
fAB = fBC + fBD (flow conservation at node B)
fBC + fDC = fCE (flow conservation at node C)
Advanced linear optimization  239
fAD + fBD = fDC + fDE (flow conservation at node D)
fCE + fDE = x (flow conservation at node E)
fAB  10, fCE  10, fDC  10
fAD  5, fBC  5, fBD  5, fDE  5
x 0
fij  0, 8(i, j) 2 A. ‍
Example 16.11: This is a shortest path problem. Figure 16.4 shows a road transpor-
tation network. The numbers on the arrows represent the travel time of each road.
Develop a linear optimization model to calculate the shortest travel time from node
A to node E. You are not allowed to enumerate all paths from A to E.
Solution. To address the problem, consider transporting one unit of cargo from
A to E. The capacity of each road is infinity, and the unit transportation cost on a
road is equal to its travel time. Then, in the optimal solution to the correspond-
ing minimum cost flow problem, the cargo will travel on the shortest path. Let
‍N := fA, B, C, D, E, F, Gg‍be the set of nodes and ‍A ‍be the set of arcs:
A := f(A, B), (A, D), (A, F), (B, C), (B, D), (B, F), (B, G),
‍ (C, E), (C, G), (D, C), (D, E), (F, G), (G, E)g.‍
Let ‍fij ‍be the decision variables representing the flow on arcs ‍(i, j) 2 A .‍ The model
is as follows:
min 12fAB + 4fAD + 3fAF + 6fBC + 2fBD + fBF + 11fBG +
5fCE + 10fCG + 23fDC + 7fDE + 8fFG + 9fGE
‍ ‍
subject to
fAB + fAD + fAF = 1 (flow conservation at A)
fAB = fBC + fBD + fBF + fBG (flow conservation at B)
‍ fBC + fDC = fCE + fCG (flow conservation at C)
fAD + fBD = fDC + fDE (flow conservation at D)
fCE + fDE + fGE = 1 (flow conservation at E)
fAF + fBF = fFG (flow conservation at F)
fBG + fCG + fFG = fGE (flow conservation at G )
fij  0, 8(i, j) 2 A.

Figure 16.4   A road transportation network

 fAB + fAD + fAF = 1 (flow conservation at A)
fAB = fBC + fBD + fBF + fBG (flow conservation at B)
240  Machine learning and data analytics for maritime studies
fBC + fDC = fCE + fCG (flow conservation at C)
fAD + fBD = fDC + fDE (flow conservation at D)
fCE + fDE + fGE = 1 (flow conservation at E)
fAF + fBF = fFG (flow conservation at F)
fBG + fCG + fFG = fGE (flow conservation at G )
fij  0, 8(i, j) 2 A. ‍
The arcs with positive flows form the shortest path.
It is possible that if there is more than one shortest path in Example 16.11, a
minimum cost flow solution may contain several paths. However, all of them are
the shortest paths, and it is not difficult to identify one of them. See the example in
Figure 16.5.
In the minimum cost flow problem in Example 16.8, we have only one type
of cargo (commodity). If we have several commodities that are differentiated by
their origins and destinations, we then have the multi-­commodity flow problem. The
models are similar to that of the minimum cost flow problem, except that we need to
differentiate the commodities.
Example 16.12: Consider Example 16.8. Node A is the port of Los Angeles,
node D is Atlanta, and node E is Chicago. A total of 12 units of clothes are to be
transported from A to E and 2 units of machines are to be transported from D to E.
Develop a linear optimization model to find the minimum cost of transporting all
the cargoes.
Solution. Let ‍N := fA, B, C, D, Eg‍be the set of nodes and ‍A ‍be the set of arcs:

‍ A := f(A, B), (A, D), (B, C), (B, D), (C, E), (D, C), (D, E)g.‍

Figure 16.5  A road transportation network with more than one shortest path
from A to E
 Advanced linear optimization  241

Let ‍fAE
ij ‍ be the decision variables representing the flow of commodities with origin
A and destination E (i.e., clothes) on arcs ‍(i, j) 2 A ‍. Let ‍fDE
ij ‍ be the decision vari-
ables representing the flow of commodities with origin D and destination E (i.e.,
machines) on arcs ‍(i, j) 2 A ‍. The model is as follows:
AE + fDE ) + 5(fAE + fDE ) + 8(fAE + fDE )]+
min 2  [6(fAB AB CE CE DC DC
AE + fDE ) + 6(fAE + fDE ) + 2(fAE + fDE ) + 7(fAE + fDE )]
3  [4(fAD
‍ AD BC BC BD BD DE DE ‍ 

subject to
Flow conservation equations:
AE + fAE = 12
fAB AD
AE = fAE + fAE
fAB BC BD
AE + fAE = fAE
fBC DC CE
AE + fAE = fAE + fAE
fAD BD DC DE
fAE AE
CE + fDE = 12
DE + fDE = 0
fAB AD
DE = fDE + fDE
fAB BC BD
DE + fDE = fDE
fBC DC CE
DE + fDE )  (fDE + fDE ) = 2
(fDC DE AD BD
fDE DE
CE + fDE = 2
Capacity constraints:
AE + fDE  10, fAE + fDE  10, fAE + fDE  10
fAB AB CE CE DC DC
AE + fDE  5, fAE + fDE  5, fAE + fDE  5, fAE + fDE  5
fAD AD BC BC BD BD DE DE
Nonnegativity constraints:

‍ fAE
ij  0, fDE
ij  0, 8(i, j) 2 A.‍

Example 16.13: Figure 16.6 shows a liner container shipping network. The circles
represent ports. There are three routes r1, r2, and r3. The capacity of a ship deployed
on route r1 is E1 (TEUs). E2 and E3 have similar meanings. The volumes of contain-
ers between different OD pairs are shown in the figure. For example, is the demand
(TEUs/week) from port 1 to port 2. The profit for transporting one TEU from port
1 to port 2 is ($/TEU) and have similar meanings. Assume that container handling
costs are 0. Develop a linear optimization model for finding the maximum profit that
can be gained from transporting containers.
It should be noted that container liner shipping networks are usually very sparse.
For example, if we assume that there is at most one arc from one node to another,
then a network with 100 nodes has at most 100 ×99 = 9 900links. If the number of
242  Machine learning and data analytics for maritime studies

Figure 16.6  A container liner shipping network with direct delivery and
transshipment

links is much smaller than this maximum number, then we say that the network is
sparse. A sparse network does not have many paths from one node to another, and
hence we can enumerate all paths*.
Solution. The set of OD pairs is ‍W = f(1, 2), (1, 3), (3, 2)g.‍ Let ‍ Hod ‍be the set of
itineraries (paths) for OD pair ‍(o, d) 2 W .‍ To simplify the notation, we use <r,‍ i >‍to
represent leg i‍ ‍of route ‍r ‍. Then ‍ H12‍consists of the following:
h1 :< 1, 1 >
‍ h2 :< 3, 1 >!< 2, 2 >‍
‍H13‍consists of:
h3 :< 3, 1 >
‍ h4 :< 1, 1 >!< 2, 1 >‍
‍H32‍consists of:
h5 :< 2, 2 >
‍ h6 :< 3, 2 >!< 1, 1 > .‍
Define ‍H := [(o,d)2W Hod .‍ Let yh be the decision variables representing the flow on
itinerary ‍h 2 H‍(TEUs/week). The model is as follows:
P od P
max g yh (maximize the total profit)
‍ (o,d)2W h2Hod ‍
subject to
P
yh  qod , (o, d) 2 W (limited demand)
h2Hod

‍ y1 + y4 + y6  E1 (capacity on leg < 1, 1 >)

y4  E2 (capacity on leg < 2, 1 >)
y2 + yof
Another advantage
* 
5  E2path
using (capacity on leg < over
flow formulations 2, 2link
>) flow formulations is that it is easy for the
former to incorporate business considerations (cf Reference 1).
y2 + y3  E3 (capacity on leg < 3, 1 >)
y6  E3 (capacity on leg < 3, 2 >)
yh  0, h 2 H.
 P
yh  qod , (o, d) 2 W (limited demand)
h2Hod
Advanced linear optimization  243
y1 + y4 + y6  E1 (capacity on leg < 1, 1 >)
y4  E2 (capacity on leg < 2, 1 >)
y2 + y5  E2 (capacity on leg < 2, 2 >)
y2 + y3  E3 (capacity on leg < 3, 1 >)
y6  E3 (capacity on leg < 3, 2 >)
yh  0, h 2 H. ‍
Example 16.14: Using the notation similar to Example 16.13, develop a linear opti-
mization model to find the maximum profit that can be gained from transporting
containers in Figure 16.7.
Solution. The set of OD pairs is ‍W = f(1, 2), (1, 3), (2, 3)g‍. Let ‍ Hod ‍be the set of
itineraries for OD pair ‍(o, d) 2 W ‍. Then ‍ H12‍consists of the following:
h1 :< 1, 1 >
‍ h2 :< 1, 4 >‍
H13
‍ ‍consists of:
h3 :< 1, 1 >!< 1, 2 >
‍ h4 :< 1, 4 >!< 1, 2 >‍
‍H23‍consists of the following:
‍ h5 :< 1, 2 > .‍
Define ‍H := [(o,d)2W Hod ‍. Let yh be the decision variables representing the flow on
itinerary‍h 2 H‍. The model is as follows:
P od P
max g yh
‍ (o,d)2W h2Hod ‍
subject to
P
h2Hod yh  q , (o, d) 2 W
od

y1 + y3  E1 (capacity on leg < 1, 1 >)

y3 + y4 + y5  E1 (capacity on leg < 1, 2 >)
y2 + y4  E1 (capacity on leg < 1, 4 >)
‍ yh  0, h 2 H. ‍

Figure 16.7   A route with two identical legs

244  Machine learning and data analytics for maritime studies

Example 16.15. Similar to Example 16.13, develop a linear optimization model

to find the maximum profit that can be gained from transporting containers in
Figure 16.8.
Solution. The set of OD pairs is ‍W = f(1, 4), (1, 6), (2, 5), (2, 6), (3, 4), (3, 5), (3, 6)g.‍
Let ‍ Hod ‍ be the set of itineraries for OD pair ‍(o, d) 2 W .‍ Then ‍H14‍consists of the
following:
‍ h1 :< 1, 1 >!< 3, 1 >‍
‍H16‍consists of the following:
‍ h2 :< 1, 1 >!< 3, 1 >!< 5, 1 >‍
‍H25‍consists of the following:
‍ h3 :< 2, 1 >!< 3, 1 >!< 4, 1 >‍
‍H26‍consists of the following:
‍ h4 :< 2, 1 >!< 3, 1 >!< 5, 1 >‍
‍H34‍consists of the following:
‍ h5 :< 3, 1 >‍
‍H35‍consists of the following:
‍ h6 :< 3, 1 >!< 4, 1 >‍
‍H36‍consists of the following:

‍ h7 :< 3, 1 >!< 5, 1 > .‍

Define ‍H := [(o,d)2W Hod .‍ Let yh be the decision variable representing the flow on
itinerary ‍h 2 H‍. The model is as follows:
P P
‍ max (o,d)2W god h2Hod yh‍ 

Figure 16.8   A hub-­and-­spoke liner shipping network

 Advanced linear optimization  245

subject to
P
yh  qod , (o, d) 2 W
h2Hod
y1 + y2  E1 (capacity on leg < 1, 1 >)
y3 + y4  E2 (capacity on leg < 2, 1 >)
y1 + y2 + y3 + y4 + y5 + y6 + y7  E3 (capacity on leg < 3, 1 >)
y3 + y6  E4 (capacity on leg < 4, 1 >)
y2 + y4 + y7  E5 (capacity on leg < 5, 1 >)
‍ yh  0, h 2 H. ‍
We encapsulate this sub-­section by stating the problems discussed in mathemati-
cal languages. We are given a network denoted by g ‍= (N , A)‍, where ‍ N ‍is the
set of nodes and ‍A  N  N ‍ is the set of arcs. (i) In the maximum flow prob-
lem, each arc ‍(i, j) 2 A ‍ has a capacity denoted by ‍si,j  0,‍ and the objective is to
find the maximum flow from node ‍nsource 2 N ‍ to node ‍nsink 2 N / fnsource g‍. (ii) In
the shortest path problem, each arc ‍(i, j) 2 A ‍ has a unit cost denoted by ‍cij  0,‍
and the objective is to find the minimum cost path from node‍nsource 2 N ‍ to node
‍n 2 N / fn
sink source
g‍. (iii) In the minimum cost flow problem, each arc ‍(i, j) 2 A ‍ has
a unit cost denoted by ‍cij  0‍ and a capacity denoted by ‍sij  0,‍ and the objec-
tive is to find the minimum total cost for transporting units of cargoes from node
‍nsource 2 N ‍ to node ‍nsink 2 N / fnsource g‍. (iv) In the multi-­commodity flow problem†,
each arc ‍(i, j) 2 A ‍ has a unit cost denoted by ‍cij  0‍ and a capacity denoted by
‍sij  0.‍ There are a set of OD pairs denoted by‍W  N  N ‍. Cargoes between
different OD pairs are different. Between OD pair ‍(o, d) 2 W, qod ‍units of cargoes
must be transported. The objective is to find the minimum total cost for transport-
ing all cargoes between all the OD pairs in ‍W ‍ ‡.

16.2 Dummy nodes and links

We have studied the maximum flow problem, the shortest path problem, the mini-
mum cost flow problem, and the multi-­commodity flow problem. Some new prob-
lems look different but can be transformed into these known problems after some
modifications. As a result, we can use approaches that we are familiar with to model
the new problems.

† 
More linear optimization formulations for the multi-­commodity flow problem can be found in Refer-
ence 2.
‡ 
In fact, the maximum flow problem and the shortest path problem are special cases of the minimum
cost flow problem; the minimum cost flow problem is a special case of the multi-­commodity flow
problem.
246  Machine learning and data analytics for maritime studies

Example 16.16: Reconsider Example 16.1. How do you add dummy§ nodes and
links to transform the problem to a minimum cost flow problem?
Solution. See Figure 16.9. The nodes W1–W3 represent the three warehouses,
and S1–S5 represent the five supermarkets. The links from a warehouse to a super-
market represent the transportation of products. The unit cost of such a link is equal
to the transportation cost, and the capacity is infinite. We further add a dummy
source node that is connected to each warehouse. A dummy arc from the dummy
source node to a warehouse has no cost and a capacity that is equal to the number of
products available in the warehouse. A dummy sink node is added and is connected
to each supermarket. A dummy arc from a supermarket to the dummy sink node has
no cost and a capacity that is equal to the number of products required by the super-
market. The problem now becomes: how to transport 8‍ 0 + 90 + 70 + 60 + 50 = 350‍
products from the dummy source node to the dummy sink node at the lowest total
cost.
Example 16.17: Reconsider Example 16.12. How do you add dummy nodes
and links to transform the problem into a minimum cost flow problem?
Solution. See Figure 16.10. The transportation network is not changed: the unit
cost and capacity of rail and truck links are the same as their physical meanings. A
dummy source node is added and is connected to nodes A and D by two dummy arcs.
The dummy arc from the dummy source node to node A has no cost and a capacity of
12; the dummy arc from the dummy source node to node D has no cost and a capac-
ity of 2. A dummy sink node is added and is connected by a dummy arc from node E.
The dummy arc has no cost and infinite capacity. The problem now becomes: how to

Figure 16.9   A transportation problem with dummy nodes and links

The word “dummy” means something that does not exist or is imaginary. For example, we can almost
§ 

always add a dummy arc to a network while imposing that its capacity is 0.
 Advanced linear optimization  247

Figure 16.10   A multi-­commodity flow problem with only one destination

transport ‍12 + 2 = 14‍units of cargoes from the dummy source node to the dummy sink
node at minimum cost.
It should be noted that this example is a special multi-­commodity flow problem
in that all commodities have the same destination. Therefore, we can transform it
into a minimum cost flow problem. If all commodities have the same origin, we can
also transform the problem into a minimum cost flow problem. However, a general
multi-­commodity flow problem cannot be transformed into a minimum cost flow
problem.
Because of trade imbalance, for example, China exports more cargoes to the
United States than the imported cargoes from the United States, some ports (e.g.,
ports in the United States) have surplus empty containers, and some ports (e.g., ports
in China) are in shortage of empty containers (these ports are called deficit ports).
Hence, empty containers have to be repositioned from surplus ports to the deficit
ports. Unlike laden containers, all empty containers can be considered identical.
Example 16.18: This is an empty container repositioning problem. Reconsider
Example 16.13. Suppose that the remaining capacity of leg <r,‍ i >‍ after carrying
laden containers is ‍EQ ri ‍. The remaining capacity is used for repositioning empty con-
tainers. The number of surplus empty containers at port ‍p 2 f1, 2, 3g‍ is denoted by
‍qQp.‍ Here ‍qQ 1 < 0,‍ ‍qQ 2 > 0,‍ and ‍qQ 3 > 0,‍P
which means port 1 is a deficit port and ports 2
3
and 3 are surplus ports. Moreover, ‍ p=1 qQp = 0.‍ We need to check whether the net-
work has sufficient capacity to reposition all the empty containers from ports 2 and 3
to port 1. How to transform the problem into a maximum flow problem?
Solution. Since all empty containers are identical, we can add a dummy source
node, connected to each surplus port with an arc whose capacity equals the volume
of the surplus empty containers at the port, and a dummy sink node, connected from
each deficit port with an arc whose capacity equals the volume of the deficit empty
containers at the port, as shown in Figure 16.11. Now the problem is transformed
into a maximum flow problem: all empty containers can be repositioned if and only
if the maximum flow from the source to the sink equals ‍qQ 2 + qQ 3.‍
Example P316.19: Reconsider Example 16.18. Suppose that ‍qQ 1 < 0,‍ ‍qQ 2 > 0,‍
Q‍ 3 < 0,‍ and ‍ p=1
q qQp = 0.‍ How do you add dummy nodes and links to transform the
empty container repositioning problem into a maximum flow problem?
Solution. See Figure 16.12. Now the problem is transformed into a maximum
flow problem: all empty containers can be repositioned if and only if the maximum
flow from the source to the sink equals ‍qQ 2.‍
248  Machine learning and data analytics for maritime studies

Figure 16.11  A liner shipping network for empty container repositioning with
dummy nodes and links

Figure 16.12  A liner shipping network for empty container repositioning with
dummy nodes and links

16.3 Using linear formulations for nonlinear problems

Some seemingly nonlinear problems can be transformed into linear problems. For
example, the constraint |x‍ + 2y|  3‍ is equivalent to the combination of ‍x + 2y  3‍
and ‍x + 2y  3;‍ the constraint ‍maxfx  2y, 2x + z, x + y  zg  z + 1‍is equivalent to
 Advanced linear optimization  249

the combination of ‍x  2y  z + 1,‍ ‍2x + z  z + 1,‍ and ‍x + y  z  z + 1.‍¶ Minimizing

the objective function ‍x + 2y + maxfx  y, 2x + 3yg‍ is equivalent to minimizing
‍x + 2y + u,‍ subject to ‍u  x  y ‍ and ‍u  2x + 3y .‍ Transforming a nonlinear con-
straint or a nonlinear model to a linear one is called linearization of the constraint
or model.
Example 16.20: We have a set of vectors ‍‍defined by all of the vectors ‍(x, y, z)‍
satisfying:
x + 2y  6
xyz4
3x + 2z  5
0x5
0y5
‍ 0  z  5.‍ 
Formulate a linear optimization model to find the vector ‍(x, y, z) 2 ‍with the small-
est L1 distance to point ‍(1, 1, 1)‍, whose definition is |x ‍  1| + |y  1| + |z  1|‍.
Solution. The problem is to minimize |x ‍  1| + |y  1| + |z  1|‍ such that
‍(x, y, z) 2 ‍. We introduce three auxiliary variables u1, u2, and u3. The model is then
equivalent to minimizing ‍u1 + u2 + u3‍such that ‍u1  x  1‍, ‍u1  1  x ‍, ‍u2  y  1‍,
‍u2  1  y ‍, ‍u3  z  1‍, ‍u3  1  z ‍, and ‍(x, y, z) 2 ‍.
Some nonlinear optimization models can be approximated by linear optimization
models. For example, if we minimize ‍x2 + y‍, and we know that 1‍  x  3‍, then we
might use the maximum of the three tangent lines at the points ‍(1, 1)‍, ‍(2, 4)‍, and ‍(3, 9)‍
in Figure 16.13 to approximate ‍x2‍. The objective function will be approximated by
‍maxf2x  1, 4x  4, 9x  18g + y‍. It should be noted that we will slightly underes-
timate the objective function, and the optimal solution may also be slightly changed
due to the change in the objective function. Nevertheless, if we use a large number of
tangent lines, the error caused by approximation will be very small**.
If a constraint is nonlinear, then we might use secant lines to approximate the
curve. For example, if we have a constraint ‍x2 + y  3‍, and we know that 1‍  x  3‍,
then we might use the maximum of the two secant lines connecting ‍(1, 1)‍ and ‍(2, 4)‍,
and ‍(2, 4)‍and ‍(3, 9)‍in Figure 16.14 to approximate ‍x2‍. The constraint will be approxi-
mated by ‍maxf3x  2, 5x  6g + y  3‍. It should be noted that some feasible solutions
will be infeasible due to the approximation. If we had used tangent lines, then some
infeasible solutions would be feasible and it would be possible that the optimal solution
to the approximation model is infeasible to the original nonlinear model.
It may also be necessary to use tangent planes. For example, we minimize
‍x ln x + y ln y  (x + y) ln(x + y)‍ with ‍x > 0, y > 0‍. The term ‍x ln x ‍ is convex in ‍x ‍. If
we let ‍z = x + y ‍and replace the objective function by ‍x ln x + y ln y  z ln z ‍, then this

The notation “max” followed by a set means the element in the set with the largest value. Some people
¶ 

use max(x, y) instead of max{x, y}.

** 
An application can be found in Reference 3.
250  Machine learning and data analytics for maritime studies

Figure 16.13   Using tangent lines to approximate a nonlinear objective function

function is no longer convex. However, we can easily check that the Hessian of
‍x ln x + y ln y  (x + y) ln(x + y)‍is positive semi-­definite. Hence, it is convex in ‍x ‍and
‍y ‍and can be linearized using tangent planes.
Finally, some nonlinear constraints cannot be approximated by linear constraints
with high accuracy, e.g., ‍x1  sin x2‍and ‍x1  x22,‍ because they are non-­convex.

16.4 Practice

Example 16.21: Consider the following service with the port rotation:
Shanghai(S, 1) ! HongKong(H, 2) ! Singapore(P, 3)
‍ ! Rotterdam(R, 4) ! Shanghai(S, 1)‍
Ships with a capacity of 6 000 (TEUs) are deployed to provide a weekly frequency.
The container shipment demand (TEUs/week) is as follows: Shanghai to Rotterdam
SR HR
‍q = 3500,‍ Hong Kong to Rotterdam ‍q = 2500,‍ Singapore to Rotterdam
PR RS
‍q = 1500,‍ and Rotterdam to Shanghai ‍q = 2500.‍ The profit of transporting one
 Advanced linear optimization  251

Figure 16.14   Using secant lines to approximate a nonlinear constraint

TEU ($/TEU) from Shanghai to Rotterdam ‍gSR = 2500‍, Hong Kong to Rotterdam
HR PR
‍g = 2300‍, Singapore to Rotterdam ‍g = 2000‍, and Rotterdam to Shanghai
gRS = 1500. Develop an optimization model to evaluate the maximum profit ($/
‍ ‍
week) the company can make.
Solution. Let ‍ySR‍, ‍yHR‍, ‍yPR‍, and ‍yRS ‍be the decision variables representing the
volumes of containers transported from S to R, H to R, P to R, and R to S, respec-
tively. The model is as follows:
max 2500y SR + 2300yHR + 2000yPR + 1500yRS ‍ 
‍
subject to
ySR  6000 (ship capacity constraint on the leg from S to H)
‍ ySR + yHR  6000 (ship capacity constraint on the leg from H to P)
ySR + yHR + yPR  6000 (ship capacity constraint on the leg from P to R)
yRS  6000 (ship capacity constraint on the leg from R to S)
0  ySR  3500
0  yHR  2500
0  yPR  1500
0  yRS  2500.
 ySR  6000 (ship capacity constraint on the leg from S to H)
252  Machine learning and data analytics for maritime studies
ySR + yHR  6000 (ship capacity constraint on the leg from H to P)
ySR + yHR + yPR  6000 (ship capacity constraint on the leg from P to R)
yRS  6000 (ship capacity constraint on the leg from R to S)
0  ySR  3500
0  yHR  2500
0  yPR  1500
0  yRS  2500.‍
Note: It is also correct if you only have the following constraints because the other
three constraints are implied by them:

ySR + yHR + yPR  6000 (ship capacity constraint on the leg from P to R)
0  ySR  3500
0  yHR  2500
0  yPR  1500
‍ 0  yRS  2500. ‍

Example 16.22: Consider the following service with the port rotation:

Shanghai (S, 1) ! Hong Kong (H, 2) ! Singapore (P, 3)

‍ ! Rotterdam (R, 4) ! Shanghai (S, 1)‍

Ships with a capacity of 6000 (TEUs) are deployed to provide a weekly fre-
quency. The container shipment demand (TEUs/week) is as follows: Shanghai
to Rotterdam ‍qSR = 3500,‍ Hong Kong to Rotterdam ‍qHR = 2500,‍ Singapore to
Rotterdam ‍qPR = 1500,‍ Rotterdam to Shanghai ‍qRS = 2500,‍ Rotterdam to Hong
Kong ‍qRH = 1500,‍ Rotterdam to Singapore ‍qRP = 500,‍ and Shanghai to Singapore
SP
‍q = 1400.‍ The profit of transporting one TEU ($/TEU) from Shanghai to
Rotterdam ‍gSR = 2500,‍ Hong Kong to Rotterdam ‍gHR = 2300,‍ Singapore to
Rotterdam ‍gPR = 2000,‍ Rotterdam to Shanghai ‍gRS = 1500,‍ Rotterdam to Hong
Kong ‍gRH = 1500,‍ Rotterdam to Singapore ‍gRP = 1500,‍ and Shanghai to Singapore
SP
‍g = 1000.‍ Develop an optimization model to evaluate the maximum profit ($/
week) the company can make.
Solution. Let ‍ySR‍, ‍yHR‍, ‍yPR‍, ‍yRS ‍, ‍yRH ‍, ‍yRP ‍, and ‍ySP ‍ be the decision variables
representing the volumes of containers transported from S to R, H to R, P to R, R to
S, R to H, R to P, and S to P, respectively. The model is as follows:
SR HR PR RS RH RP SP
‍max 2500y + 2300y + 2000y + 1500y + 1500y + 1500y + 1000y ‍ 

subject to
ySR + yRH + yRP + ySP  6000 (ship capacity constraint on the leg from S to H)
ySR + yHR + yRP + ySP  6000 (ship capacity constraint on the leg from H to P)
‍ ySR + yHR + yPR  6000 (ship capacity constraint on the leg from P to R)
yRS + yRH + yRP  6000 (ship capacity constraint on the leg from R to S)
0  ySR  3500
0  yHR  2500
0  yPR  1500
0  yRS  2500
0  yRH  1500
 ySR + yRH + yRP + ySP  6000 (ship capacity constraint on the leg from S to H)
ySR + yHR + yRP + ySP  6000 (ship capacity constraint on the leg from H to P)
Advanced linear optimization  253
ySR + yHR + yPR  6000 (ship capacity constraint on the leg from P to R)
yRS + yRH + yRP  6000 (ship capacity constraint on the leg from R to S)
0  ySR  3500
0  yHR  2500
0  yPR  1500
0  yRS  2500
0  yRH  1500
0  yRP  500
0  ySP  1400.‍ 

Example 16.23: Consider the following service with the port rotation:

Shanghai(S, 1) ! HongKong(H, 2) ! Singapore(P, 3)

‍ ! Rotterdam(R, 4) ! HongKong(H, 5) ! Shanghai(S, 1)‍

Ships with a capacity of 6000 (TEUs) are deployed to provide a weekly frequency.
The container shipment demand (TEUs/week) is as follows: Shanghai to Rotterdam
qSR = 3500, Hong Kong to Rotterdam qHR = 2500, Singapore to Rotterdam qPR =
1500, Rotterdam to Shanghai qRS = 2500, Rotterdam to Hong Kong qRH = 1500,
Rotterdam to Singapore qRP = 500, and Shanghai to Singapore qSP = 1400. The profit
of transporting one TEU ($/TEU) from Shanghai to Rotterdam gSR = 2500, Hong
Kong to Rotterdam gHR = 2300, Singapore to Rotterdam gPR = 2000, Rotterdam to
Shanghai gRS = 1500, Rotterdam to Hong Kong gRH = 1500, Rotterdam to Singapore
gRP = 1500, and Shanghai to Singapore gSP = 1000.
(i) Develop an optimization model to evaluate the maximum profit ($/week) the
company can make.
(ii) Write down two distinct feasible solutions, and calculate their objective
function values.
(iii) Is this model infeasible? Why? Is this model unbounded? Why? Does this
model have an optimal solution? Why?
Solution. (i) Let ‍ySR‍, ‍yHR‍, ‍yPR‍, ‍yRS ‍, ‍yRH ‍, ‍yRP ‍, and ‍ySP ‍be the decision variables
representing the volumes of containers transported from S to R, H to R, P to R, R to
S, R to H, R to P, and S to P, respectively. The model is as follows:
SR HR PR RS RH RP SP
‍max 2500y + 2300y + 2000y + 1500y + 1500y + 1500y + 1000y ‍ 
subject to
ySR + yRP + ySP  6000 (ship capacity constraint on the leg from S to H)
ySR + yHR + yRP + ySP  6000 (ship capacity constraint on the leg from H to P)
ySR + yHR + yPR  6000 (ship capacity constraint on the leg from P to R)
yRS + yRH + yRP  6000 (ship capacity constraint on the leg from R to H)
‍ yRS + yRP  6000 (ship capacity constraint on the leg from H to S)
0  ySR  3500
0  yHR  2500
0  yPR  1500
0  yRS  2500
0  yRH  1500
RP
 ySR + yRP + ySP  6000 (ship capacity constraint on the leg from S to H)
ySR + yHR + yRP + ySP  6000 (ship capacity constraint on the leg from H to P)
ySR + yHR + yPR  6000 (ship capacity constraint on the leg from P to R)
yRS + yRH + yRP  6000 (ship capacity constraint on the leg from R to H)
254  Machine learning and data analytics for maritime studies
yRS + yRP  6000 (ship capacity constraint on the leg from H to S)
0  ySR  3500
0  yHR  2500
0  yPR  1500
0  yRS  2500
0  yRH  1500
0  yRP  500
0  ySP  1400.‍ 

(ii) One feasible solution is, and its objective function value is 0. Another feasible
solution is, and its objective function value is 2 500. (iii) The model is feasible,
because we have found two feasible solutions in (ii). The feasible set of the model
is bounded, because letting, any feasible solution satisfies the following. Hence, the
model is not unbounded. Since this is a linear optimization model that is feasible and
not unbounded, it has an optimal solution.
Example 16.24: Reconsider the CC1 service with the port rotation:
Shanghai (S, 1) ! Kwangyang (K, 2) ! Pusan (P, 3) ! Los Angeles (L, 4)
‍ ! Oakland(O, 5) ! Pusan (P, 6) ! Kwangyang(K, 7) ! Shanghai (S, 1) ‍ 
Ships with a capacity of (TEUs) are deployed to provide a weekly frequency. The
container shipment demand (TEUs/week) is as follows: Shanghai to Los Angeles
qSL, Pusan to Los Angeles qPL, Oakland to Shanghai qOS, and Shanghai to Oakland
qSO. The profit of transporting one TEU ($/TEU) from Shanghai to Los Angeles is
gSL, Pusan to Los Angeles gPL, Oakland to Shanghai gOS, and Shanghai to Oakland
gSO. The values of E, q’s, and g’s are all given.
(i) Develop an optimization model to evaluate the maximum profit ($/week) the
company can make.
(ii) If qSL is increased, how will the feasible set of the model in (i) change, and
how will the optimal objective function value change?
(iii) If E is increased, how will the feasible set of the model in (i) change, and
how will the optimal objective function value change?
(iv) If the values of gSL, gPL, gOS, gSO and are all doubled simultaneously, how
will the feasible set of the model in (i) change, and how will the optimal objective
function value change?
Solution. (i) Define the set of OD pairs W ‍ = f(S, L), (P, L), (O, S), (S, O)g‍. Let
od
‍y ‍ be the decision variables representing the volumes of containers transported
between OD pairs ‍(o, d) 2 W ‍. The model is as follows:
P od od
max g y
‍ (o,d)2W ‍
subject to
ySL + ySO  E (ship capacity constraint on the leg from S to K)
‍ ySL + ySO  E (ship capacity constraint on the leg from K to P)
ySL + yPL + ySO  E (ship capacity constraint on the leg from P to L)
ySO  E (ship capacity constraint on the leg from L to O)
yOS  E (ship capacity constraint on the leg from O to P)
yOS  E (ship capacity constraint on the leg from P to K)
yOS  E (ship capacity constraint on the leg from K to S)
0  yod  qod , (o, d) 2 W.
 ySL + ySO  E (ship capacity constraint on the leg from S to K)
Advanced linear optimization  255
ySL + ySO  E (ship capacity constraint on the leg from K to P)
ySL + yPL + ySO  E (ship capacity constraint on the leg from P to L)
ySO  E (ship capacity constraint on the leg from L to O)
yOS  E (ship capacity constraint on the leg from O to P)
yOS  E (ship capacity constraint on the leg from P to K)
yOS  E (ship capacity constraint on the leg from K to S)
0  yod  qod , (o, d) 2 W.‍ 

Note: It is also correct if you only have the following constraints because the other
five constraints are implied by them:

ySL + yPL + ySO  E (ship capacity constraint on the leg from P to L)

yOS  E (ship capacity constraint on the leg from K to S)
‍ 0  yod  qod , (o, d) 2 W.‍ 

(ii) The feasible set will be larger or unchanged; the optimal objective function value
will be larger or unchanged.
(iii) The feasible set will be larger or unchanged; the optimal objective function
value will be larger or unchanged.
(iv) The feasible set will be unchanged; the optimal objective function value
will be doubled.
Example 16.25: (i) Write down an example of a linear optimization model with
two variables and three constraints that is infeasible.
(ii) Write down an example of a linear optimization model with two variables
and three constraints that is unbounded.
(iii) Write down an example of a linear optimization model with two variables
and three constraints that has an infinite number of optimal solutions.
(iv) Write down an example of a linear optimization model with two variables
and three constraints whose optimal objective function value is  ‍ 123‍.
Solution. Note: The answers are not unique.
(i) Minimize subject to x + y subject to x ≤ 0, x ≥2, y ≥ o.
(ii) Minimize subject to x + y subject to x ≤ 0, x≤ −2, y ≥ 0.
(iii) Minimize subject to x + y subject to x + y ≥ 1, x ≥ 0, y ≥ 0.
(iv) Minimize subject to x + y subject to x ≥ − 123, x ≥ −200, y ≥ 0.
Example16.26: We have the following linear optimization model:
‍ min x  y ‍ 
subject to
x+y3
y2
‍ x  2y = 1.‍ 
(i) Transform the model to the canonical form that maximizes ‍cT x ‍subject to ‍Ax  b‍
and ‍x  0‍.
256  Machine learning and data analytics for maritime studies

(ii) Transform the model to the standard form that maximizes ‍cT x ‍ subject to
‍Ax = b‍and ‍x  0.‍
Note: The answers are not unique.
Solution. Letting ‍x = u  v, u  0, v  0‍ and ‍w = 2  y, w  0,‍ the model is
transformed to
‍ min(u  v)  (2  w)‍ 
subject to
(u  v) + (2  w)  3
(u  v)  2(2  w) = 1
u0
v0
‍ w  0.‍ 
(i) The model is equivalent to
‍ max (u  v) + (2  w)‍ 
subject to
(u  v)  (2  w)  3
(u  v)  2(2  w)  1
(u  v) + 2(2  w)  1
u0
v0
‍ w  0‍ 
which is equivalent to
‍ max u + v  w ‍ 
subject to
u + v + w  1
u  v + 2w  5
u + v  2w  5
u0
v0
‍ w  0.‍ 
2 3
1 1 1
6 7
6 27
h i A =4 1 1 5
x= u h i
v w T 1
1 2 ‍, b= 1
Hence, hin the canonical
i form, ‍ ‍, ‍ 5 5 T
,
‍ ‍
and c= 1 1 1 , with ‍x = u  v ‍and ‍y = 2  w ‍.
T
‍ ‍
 Advanced linear optimization  257

(ii) The model is equivalent to

‍ max (u  v) + (2  w)‍ 
subject to
(u  v) + (2  w)  t = 3
(u  v)  2(2  w) = 1
u0
v0
w0
‍ t  0‍ 
which is equivalent to
‍ max u + v  w ‍ 
subject to
uvwt= 1
u  v + 2w = 5
u0
v0
w0
‍ t  0.‍  " #
h i 1 1 1 1
A=
x= u v w t T 1 1 2 0 ‍,
Hence, in the standardh form, ‍ i ‍, ‍
‍b=[ 1 5 ] ‍, and c= 1 1 1 0 , with ‍x = u  v ‍and ‍y = 2  w .‍
T T
‍ ‍
Example 16.27: Consider the model below:
‍ max x + 2y ‍
subject to
x+y1
x+y2
x0
‍ y  0.‍ 

(i) Use the graphical method to find the optimal solution.

(ii) Propose another objective function that will make the model have an infinite
number of optimal solutions. You are not allowed to propose a function that is a
constant, e.g., minimize 1.
(iii) Write down a linear optimization model satisfying: first, it has two vari-
ables; second, its objective function is not a constant; third, its feasible set is a pen-
tagon; and fourth, it has exactly one optimal solution. Use the graphical method to
find the optimal solution.
258  Machine learning and data analytics for maritime studies

(iv) Write down a linear optimization model satisfying: first, it has two vari-
ables; second, its objective function is not a constant; third, its feasible set is a pen-
tagon; and fourth, it has an infinite number of optimal solutions. Use the graphical
method to find an optimal solution.
Solution. (i) See Figure 16.15. The optimal solution corresponds to the intersec-
tion of the lines ‍x = 0‍and ‍x + y = 2.‍ Therefore, the optimal solution is ‍x = 0, y = 2.‍
The optimal objective function value is 4.
(ii) Any of the following solutions are correct:
‍min x ‍, ‍min y ,‍ ‍min x + y ,‍ or ‍max x + y .‍
(iii) An example is
‍min x + y ‍
subject to
x2
y2
x+y3
x0
‍ y  0.‍ 

The graphical method is not provided here. The optimal solution is ‍x = 0, y = 0,‍
and the optimal objective function value is 0.
(iv) We can change the objective function in the model in (iii) to ‍min x ‍. The
graphical method is not provided here. An optimal solution is ‍x = 0, y = 0,‍and the
optimal objective function value is 0.

Figure 16.15   Graphical method

 Advanced linear optimization  259

Example 16.28: Use the graphical method to find the optimal solution to the
model below:

‍ max x + 2y ‍ 
subject to
x+y= 1
x+y2
x0
‍ y  0.‍ 

Solution. See Figure 16.16. The optimal solution corresponds to the intersection of

the lines ‍x = 0‍ and ‍x + y = 1‍. Therefore, the optimal solution is ‍x = 0, y = 1‍. The
optimal objective function value is 2.
Example 16.29: Use Excel or any other software to solve the following linear
optimization model. You only need to write down the final answer:
‍ min x + 2y + 3z ‍
subject to

4x + 5y + 6z  10000
7x + 8y + 9z  100
‍ 10x + 11y + 12z  200
13x + 14y + 15z  300
x0
y0
z  0.

Figure 16.16   Graphical method

 4x + 5y + 6z  10000
7x + 8y + 9z  100
260  Machine learning and data analytics for maritime studies
10x + 11y + 12z  200
13x + 14y + 15z  300
x0
y0
z  0.‍
Solution. The optimal solution is ‍x = k/4, y = 0, z = 0.‍ The optimal objective
function value is ‍k/4‍.
Example 16.30: We have a transportation network shown in Figure 16.17, where
the set of nodes ‍ N ‍and the set of arcs ‍A ‍are shown. Each arc ‍(i, j) 2 A ‍has a known
capacity denoted by ‍sij  0.‍ Develop a link-­flow linear optimization model to find the
maximum flow from A to F. You must P write down the details of the objective func-
tion and constraints without using “‍ ”‍ or “‍8‍” except in the nonnegativity constraints.
Solution. Let fijbe the decision variables representing the flow on arcs ‍(i, j) 2 A .‍
The model is as follows:

‍ max fAB + fAC + fAF (maximize the total outflow from node A)‍

subject to
fAB + fDB = fBC + fBE
(flow conservation at node B)
fAC + fBC = fCD + fCE
(flow conservation at node C)
fCD + fED = fDB + fDF
(flow conservation at node D)
fBE + fCE = fED + fEF
(flow conservation at node E)
fAB  sAB , fAC  sAC , fAF  sAF , fBC  sBC , fBE  sBE
fCD  sCD , fCE  sCE , fDB  sDB , fDF  sDF , fED  sED , fEF  sEF
‍ fij  0, 8(i, j) 2 A. ‍ 

Figure 16.17   A transportation network

 Advanced linear optimization  261

Example 16.31: We have a transportation network shown in Figure 16.17, where

the set of nodes ‍ N ‍and the set of arcs ‍A ‍are shown. Each arc ‍(i, j) 2 A ‍has a known
unit cost denoted by ‍cij  0‍ and a known capacity denoted by ‍sij  0.‍ We need to
send 1‍ 0‍units of cargo from A to E (note it is from A to E rather than to F). Develop a
link-­flow linear optimization model to find the minimum total cost. You must P write
down the details of the objective function and constraints without using “‍ ”‍ or “‍8
‍” except in the nonnegativity constraints.
Solution. Let fij be the decision variables representing the flow on arcs ‍(i, j) 2 A .‍
The model is as follows:
min cAB fAB + cAC fAC + cAF fAF + cBC fBC + cBE fBE
‍ +cCD fCD + cCE fCE + cDB fDB + cDF fDF + cED fED + cEF fEF ‍
subject to
fAB + fAC + fAF = 10
fAB + fDB = fBC + fBE
fAC + fBC = fCD + fCE
fCD + fED = fDB + fDF
fBE + fCE = fED + fEF + 10
fAF + fDF + fEF = 0
fAB  sAB , fAC  sAC , fAF  sAF , fBC  sBC , fBE  sBE
fCD  sCD , fCE  sCE , fDB  sDB , fDF  sDF , fED  sED , fEF  sEF
‍ fij  0, 8(i, j) 2 A. ‍ 

It is correct to say that there is no flow on (E,D), (A,F), (D,F), and (E,F). Therefore,
the following model is also correct:
min cAB fAB + cAC fAC + cBC fBC + cBE fBE +
‍ cCD fCD + cCE fCE + cDB fDB ‍

subject to
fAB + fAC = 10
fAB + fDB = fBC + fBE
fAC + fBC = fCD + fCE
fCD = fDB
fBE + fCE = 10
fAB  sAB , fAC  sAC , fBC  sBC , fBE  sBE
fCD  sCD , fCE  sCE , fDB  sDB
‍ fij  0, 8(i, j) 2 A \ f(E, D), (A, F), (D, F), (E, F)g.‍ 

Example 16.32: Figure 16.18 shows a liner container shipping network. The circles rep-
resent ports. There are four routes r1, r2, r3, and r4 that provide weekly shipping services.
262  Machine learning and data analytics for maritime studies

Figure 16.18   A complex hub and spoke liner shipping network

The capacity of a ship deployed on route r1 is E1 (TEUs). E2, E3, and E4 have similar
meanings. The volumes of containers between different OD pairs are shown in the fig-
ure. For example, ‍q15‍is the demand (TEUs/week) from port 1 to port 5. The profit for
transporting one TEU from port 1 to port 5 is ‍g15‍($/TEU). ‍g17‍, ‍g25‍, ‍g36‍, and ‍g46‍have
similar meanings. Assume that container handling costs are 0. Develop a path-­flow lin-
ear optimization model to find the maximum profit that can be gained from transporting
containers. YouPmust write down the details of the objective function and constraints
without using “‍ ‍” or “‍8‍” except in the nonnegativity constraints.
Solution. The set of OD pairs is ‍W = f(1, 5), (1, 7), (2, 5), (3, 6), (4, 6)g‍. Let ‍ Hod ‍
be the set of itineraries (paths) for OD pair ‍(o, d) 2 W ‍. To simplify the notation, we
use ‍< r, i >‍to represent leg i‍ ‍of route ‍r ‍. Then ‍ H15‍consists of the following:
‍ h1 :< 1, 2 >!< 1, 3 >!< 3, 1 >‍
‍H17‍consists of the following:

‍ h2 :< 1, 2 >!< 1, 3 >!< 3, 1 >!< 3, 2 >!< 4, 1 >‍

‍H25‍consists of the following:
‍ h3 :< 1, 3 >!< 3, 1 >‍

‍H36‍consists of the following:

‍ h4 :< 2, 1 >!< 3, 1 >!< 3, 2 >‍

‍H46‍consists of the following:

‍ h5 :< 3, 1 >!< 3, 2 > .‍
Define ‍H := f1, 2, 3, 4, 5g‍. Let yh be the decision variables representing the flow on
itinerary ‍h 2 H ‍(TEUs/week). The model is as follows:

‍ max g15 y1 + g17 (y2 ) + g25 (y3 ) + g36 (y4 ) + g46 (y5 ) (maximize the total profit)‍
 Advanced linear optimization  263

subject to
y1  q15
y2  q17
y3  q25
y4  q36
y5  q46
y1 + y2  E1 (capacity on leg < 1, 2 >)
y1 + y2 + y3  E1 (capacity on leg < 1, 3 >)
y4  E2 (capacity on leg < 2, 1 >)
y1 + y2 + y3 + y4 + y5  E3 (capacity on leg < 3, 1 >)
y2 + y4 + y5  E3 (capacity on leg < 3, 2 >)
y2  E4 (capacity on leg < 4, 1 >)
‍ yh  0, h 2 H. ‍
Example 16.33: Propose a linear optimization model that satisfies all of the fol-
lowing requirements or answer “such a model does not exist”: (i) it is infeasible,
(ii) after removing the first constraint, it has exactly one optimal solution; (iii) after
removing the first two constraints, it has an infinite number of optimal solutions; and
(iv) after removing the first three constraints, it is unbounded.
Solution. An example can be
‍ max x ‍ 
subject to
x  1
x+y1
x1
x0
‍ y  0.‍ 

Example 16.34: We have a transportation network shown in Figure 16.19, where

the set of nodes ‍ N ‍and the set of arcs ‍A ‍are shown. Each arc ‍(i, j) 2 A ‍has a capacity
denoted by ‍sij  0.‍ We do not know ‍sij ,‍ but we know the maximum flow from A to D
is 11, the maximum flow from D to G is 12, and the maximum flow from G to J is 13.
What is the maximum flow from A to J? You do not need to show the calculation.
Solution.The answer is 11.
Example 16.35: Consider a problem with three decision variables ‍x, y, z ‍ and a
set of linear constraints that defines a non-­empty and bounded feasible set. Suppose
that there are three potential objective functions to maximize: ‍x + y + z ,‍ ‍2x  y ,‍ and
‍  z .‍ How to use a linear optimization solver to check whether there is a feasible
y
solution that could maximize the three objectives at the same time?
264  Machine learning and data analytics for maritime studies

Figure 16.19   A transportation network

Solution. Solve the maximization problem with the objective function ‍x + y + z ‍

and the given constraints, and let ‍(x1 , y1 , z1 )‍ be the optimal solution. Solve the
maximization problem with the objective function ‍2x  y ‍and the given constraints,
and let ‍(x2 , y2 , z2 )‍ be the optimal solution. Solve the maximization problem with
the objective function y   
‍  z ‍ and the given constraints, and let ‍(x3 , y3 , z3 )‍ be the
optimal solution. Solve the following model: maximizing 0 subject to the given con-
straints and ‍x + y + z = x1 + y1 + z1 , 2x  y = 2x2  y2 , y  z = y3  z3 .‍ If this
model has a feasible solution, then the answer is yes; otherwise, the answer is no.
Another approach is as follows. Let C1 be the optimal objective function value
of the maximization model with the objective function ‍x + y + z ‍ and the given con-
straints, C2 be the one of the models with the objective function ‍2x  y‍and the given
constraints, C3 be the one of the models with the objective function y ‍  z ‍ and
the given constraints, and C4 be the one of the models with the objective function
‍(x + y + z) + (2x  y) + (y  z)‍and the given constraints. If C ‍ 4 = C1 + C2 + C3,‍ then
the answer is yes; otherwise, the answer is no.
Example 16.36: We have an optimization model with decision variables ‍x, y, z ,‍
linear constraints and the objective function that minimizes
‍ maxfx, y, zg  minfx, y, zg.‍

The feasible set is non-­empty. Transform the model to a linear optimization model.
Solution. Define two auxiliary decision variables ‍u‍and ‍v ‍. The model is equiva-
lent to
‍ min u  v ‍ 
subject to
u  x, u  y, u  z
‍ v  x, v  y, v  z ‍ 

and the given constraints.

Example 16.37: A linear optimization model has the following decision vari-
ables ‍x1 , x2‍and constraints:
 Advanced linear optimization  265

a11 x1 + a12 x2  b1
a21 x1 + a22 x2  b2
a31 x1 + a32 x2  b3
..
.
am1 x1 + am2 x2  bm
0  x1  1
‍ 0  x2  1.‍

Suppose that we know all points in the feasible set form a square. Given a linear
optimization solver, how to calculate the area of the square?
Example 16.38: We have a linear optimization model with an infinite number
of optimal solutions. How to find two optimal solutions with the largest L1 distance?
Example 16.39: Given a linear optimization model, how to use a linear opti-
mization solver to check whether all points in the feasible set form a line segment?
Example 16.40: Walmart has three warehouses (W1, W2, and W3) that store
the same type of product and five supermarkets (S1–S5) that need the products in a
city. The number of products available at each warehouse, the number of products
needed at each supermarket, and the transportation cost per unit product from each
warehouse to each supermarket are shown below. Develop a linear optimization
model to help Walmart make the decision of how to transport the products. You
must
P write down the details of the objective function and constraints without using
“‍ ”‍ or “‍8‍” except in the nonnegativity constraints.

Warehouse Number of products available

W1 100
W2 200
W3 50

Supermarket Number of products needed

S1 80
S2 90
S3 70
S4 60
S5 50

Unit cost ($) S1 S2 S3 S4 S5

W1 1 2 4 3 6
W2 5 2 4 4 4
W3 5 1 1 3 2
266  Machine learning and data analytics for maritime studies

Solution. Let ‍fij ‍ be the decision variables representing the number of products
transported from warehouse i‍ = 1, 2, 3‍ to supermarket ‍j = 1, 2, 3, 4, 5‍. The model is
as follows:
min f11 + 2f12 + 4f13 + 3f14 + 6f15 + 5f21 + 2f22 +

‍ 4f23 + 4f24 + 4f25 + 5f31 + f32 + f33 + 3f34 + 2f35 ‍ 

subject to
f11 + f12 + f13 + f14 + f15  100
f21 + f22 + f23 + f24 + f25  200
f31 + f32 + f33 + f34 + f35  50
f11 + f21 + f31 = 80
f12 + f22 + f32 = 90
f13 + f23 + f33 = 70
f14 + f24 + f34 = 60
f15 + f25 + f35 = 50

‍ fij  0, i = 1, 2, 3, j = 1, 2, 3, 4, 5.‍

Note: The following constraints are incorrect because evidently all fij will be 0 in the
optimal solution:
f11 + f12 + f13 + f14 + f15  100
f21 + f22 + f23 + f24 + f25  200
f31 + f32 + f33 + f34 + f35  50
f11 + f21 + f31  80
f12 + f22 + f32  90
f13 + f23 + f33  70
f14 + f24 + f34  60
f15 + f25 + f35  50

‍ fij  0, i = 1, 2, 3, j = 1, 2, 3, 4, 5.‍

Example 16.41: We have the following linear optimization model:

‍ min x  y ‍ 
subject to
x+y3
y2
‍ x  2y = 1.‍ 
Transform the model to the standard form that maximizes ‍cT x ‍subject to ‍Ax = b‍and
‍x  0.‍
Solution. The model is equivalent to
‍ max (u  v) + (2  w)‍ 
 Advanced linear optimization  267

subject to
(u  v) + (2  w)  t = 3
(u  v)  2(2  w) = 1
u0
v0
w0
‍ t  0‍ 
which is equivalent to
‍ max  u + v  w ‍
subject to
uvwt= 1
u  v + 2w = 5
u0
v0
w0
‍ t  0.‍
" #
h i 1 1 1 1
A=
x= u v w t T 1 1 2 0 ‍,
Hence,
h in
i the standard
h form, ‍ i ,
‍ ‍
y = 2  w
b= 1 5 T , and c= 1 1 1 0 T , with ‍x = u  v ‍and ‍ ‍.
‍ ‍ ‍ ‍
Note 1: The solution to this question is not unique.
Note 2: The following interesting solutions are also correct: Let
u
‍ = x + y  3  0‍ and ‍v = 2  y  0.‍ Solving this equation system, we have
‍x = u + v + 1, y = 2  v .‍ Therefore, the model is equivalent to ‍min u + 2v  1‍subject
to ‍u + 3v  3 = 1, u  0, v  0,‍ etc.
Note 3: It is incorrect to say ‍y  2‍is equivalent to ‍w = y, w  2, w  0.‍
Example 16.42: Use the graphical method to find the optimal solution to the
model below:
‍ max 2x + y ‍
subject to
x+y1
x2
y1
x0
‍ y  0.‍ 
268  Machine learning and data analytics for maritime studies

Figure 16.20   Graphical method

Solution. See Figure  16.20. The optimal solution corresponds to the intersection
of the lines ‍x = 2‍ and ‍y = 1‍. Therefore, the optimal solution is ‍x = 2, y = 1‍. The
optimal objective function value is ‍5‍.
Note: The slope of the line ‍2x + y = k ‍is ‍2‍rather than ‍2‍.

Figure 16.21   A transportation network

 Advanced linear optimization  269
Example 16.43: We have a transportation network shown in Figure  16.21,
where the set of nodes ‍ N ‍and the set of arcs ‍A ‍are shown. Each arc ‍(i, j) 2 A ‍has a
known unit transportation cost denoted by ‍cij  0‍and a known capacity denoted by
‍sij  0.‍ We need to send 1‍ 0‍ units of cargo from A to F. Develop a link-­flow linear
optimization model to find the minimum total cost. You must P write down the details
of the objective function and constraints without using “‍ ”‍ or “‍8‍” except in the
nonnegativity constraints.
Solution. Let ‍fij ‍be the decision variables representing the flow on arcs ‍(i, j) 2 A .‍
The model is as follows:
min cAB fAB + cAC fAC + cAF fAF + cBC fBC + cBE fBE +
‍ cCD fCD + cCE fCE + cDB fDB + cDF fDF + cED fED + cEF fEF ‍ 
subject to
fAB + fAC + fAF = 10
fAB + fDB = fBC + fBE
fAC + fBC = fCD + fCE
fCD + fED = fDB + fDF
fBE + fCE = fED + fEF
fAF + fDF + fEF = 10
fAB  sAB , fAC  sAC , fAF  sAF , fBC  sBC , fBE  sBE
fCD  sCD , fCE  sCE , fDB  sDB , fDF  sDF , fED  sED , fEF  sEF
‍ fij  0, 8(i, j) 2 A. ‍ 

Example 16.44: Figure 16.22 shows a liner container shipping network. The circles
represent ports. There are five routes r1, r2, r3, r4, and r5 that provide weekly shipping
services. The capacity of a ship deployed on route r1 is E1 (TEUs). E2, E3, E4, and E5
have similar meanings. The volumes of containers between different OD pairs are
shown in the figure. For example, ‍q12‍ is the demand (TEUs/week) from port 1 to
port 2. The profit for transporting one TEU from port 1 to port 2 is ‍g12‍($/TEU). ‍g14‍
and ‍g15‍have similar meanings. Assume that container handling costs are 0. Develop
a path-­flow linear optimization model to find the maximum profit that can be gained
from transporting containers. You must
P write down the details of the objective func-
tion and constraints without using “‍ ”‍ or “‍8‍” except in the nonnegativity constraints.

Figure 16.22   A complex transshipment liner shipping network

270  Machine learning and data analytics for maritime studies

Solution. The set of OD pairs is ‍W = f(1, 2), (1, 4), (1, 5)g‍. Let ‍ Hod ‍be the set of
itineraries (paths) for OD pair ‍(o, d) 2 W ‍. To simplify the notation, we use <r,‍ i >‍to
represent leg i‍ ‍of route ‍r ‍. Then ‍ H12‍consists of the following:
‍ h1 :< 1, 1 >‍

‍H14‍consists of the following:

h2 :< 1, 1 >!< 3, 1 >
‍ h3 :< 1, 1 >!< 2, 1 >!< 4, 1 >!< 5, 2 >‍

‍H15‍consists of the following:

h4 :< 1, 1 >!< 2, 1 >!< 4, 1 >
‍ h5 :< 1, 1 >!< 3, 1 >!< 5, 1 > .‍
Define ‍H := f1, 2, 3, 4g‍. Let yh be the decision variables representing the flow on
itinerary (TEUs/week). The model is as follows:
‍ max g12 y1 + g14 (y2 + y3 ) + g15 (y4 + y5 ) (maximize the total profit)‍
subject to

y1  q12
y2 + y3  q14
y4 + y5  q15
y1 + y2 + y3 + y4 + y5  E1 (capacity on leg < 1, 1 >)
y3 + y4  E2 (capacity on leg < 2, 1 >)
y2 + y5  E3 (capacity on leg < 3, 1 >)
y3 + y4  E4 (capacity on leg < 4, 1 >)
y5  E5 (capacity on leg < 5, 1 >)
y3  E5 (capacity on leg < 5, 2 >)
‍ yh  0, h 2 H.‍ 
Note: It is wrong to say ‍y4  q15 , y5  q15‍.
Example 16.45: Propose a linear optimization model that satisfies all of
the following requirements or answer “such a model does not exist”: (i) it is
infeasible; (ii) after removing the first constraint, it has an infinite number of
optimal solutions; (iii) after removing the first two constraints, it has exactly
one optimal solution; and (iv) after removing the first three constraints, it is
unbounded.
Solution. Note: The answer is not unique. An example can be
‍ max x ‍ 
subject to
 Advanced linear optimization  271

x  1
x1
x+y2
x0
‍ y  0.‍ 

Example 16.46: We have a transportation network shown in Figure 16.23, where

the set of nodes ‍ N ‍and the set of arcs ‍A ‍are shown. Each arc ‍(i, j) 2 A ‍has a capac-
ity denoted by ‍sij  0.‍ We do not know ‍sij ,‍ but we know the maximum flow from
A to D is 11, the maximum flow from D to G is 12, the maximum flow from G to
234,554
J is ‍1,000,000 ,‍ the maximum flow from A to M is 14, and the maximum flow from
M to G is 15. What is the maximum flow from A to J? You do not need to show
the calculation.
Solution.

234, 554
‍ 1, 000, 000 ‍ 

Example 16.47: A linear optimization model has decision variables ‍x1 , x2 ,    , xn‍
and constraints:

Figure 16.23   A transportation network

272  Machine learning and data analytics for maritime studies

a11 x1 + a12 x2 +    + a1n xn = b1

a21 x1 + a22 x2 +    + a2n xn  b2
a31 x1 + a32 x2 +    + a3n xn  b3
..
.
am1 x1 + am2 x2 +    + amn xn  bm
0  x1  1
0  x2  1
..
.
‍ 0  xn  1‍ 
where ‍aij , bi ‍ are all constants, i‍ = 1, 2,    , m
, m,‍ ‍j = 1, 2,    , n
, n.‍ We do not know its
objective function yet. How to use a linear optimization solver to check whether the
first constraint ‍a11 x1 + a12 x2 +    + a1n xn = b1‍is redundant?
Solution. First, we minimize ‍a11 x1 + a12 x2 +    + a1n xn‍subject to the second to
the last constraints. Second, we maximize ‍a11 x1 + a12 x2 +    + a1n xn‍ subject to the
second to the last constraints. If both models have the same optimal objective func-
tion value b1, then the first constraint is redundant. Otherwise, it is not redundant.
Example 16.48: We have an optimization model with decision variables ‍x, y, z ‍,
linear constraints, and the objective function that maximizes
‍ maxfx, y, zg  minfx, y, zg.‍
The feasible set is non-­empty. How to use a linear optimization solver to address the
problem? Hint: you can use the linear optimization solver several times.
Solution. There are six possible cases: ‍x  y  z ‍, ‍x  z  y ‍, ‍y  x  z ‍,
‍y  z  x ‍, ‍z  x  y ‍, and ‍z  y  x ‍. Taking the first case as an example, we max-
imize ‍x  z ‍ subject to all the given constraints and ‍x  y, y  z ‍. Taking the last
case as an example, we maximize ‍z  x ‍ subject to all the given constraints and
‍z  y, y  x ‍. Hence, we solve six linear optimization models. The best one of the six
optimal solutions is what we want.
Note 1: The following answer is also correct: we maximize ‍x  y ‍subject to all
the given constraints; we maximize ‍x  z ‍subject to all the given constraints; and we
do so with the objective functions ‍y  x ‍, ‍y  z ‍, ‍z  x ‍, ‍z  y ‍. Hence, we solve six lin-
ear optimization models. The best one of the six optimal solutions is what we want.
Note 2: The following answer is incorrect: solve ‍max(u  v)‍ subject to
‍u  x, u  y, u  z, v  x, v  y, v  z ‍ and all of the other constraints. It is wrong
because the problem will always be unbounded: ‍u‍can be infinity and ‍v ‍can be nega-
tive infinity.
Note 3: The following answer is incorrect: ‍u‍ is the best objective func-
tion value of the following three models: ‍max x ‍ subject to all of the constraints,
‍max y ‍ subject to all of the constraints, and ‍ max z ‍subject to all of the con-
straints. ‍v‍ is the best objective function value of the following three models:
‍min x ‍ subject to all of the constraints, ‍min y ‍ subject to all of the constraints, and
 Advanced linear optimization  273

Figure 16.24   A counter-­example

‍min z ‍subject to all of the constraints. Then we solve the model ‍min 0‍ subject to
 
‍u  u , u  x, u  y, u  z, v  v , v  x, v  y, v  z ‍ and all of the other con-
straints, whose optimal solution is also optimal to the original model.
It is wrong because there may not exist a feasible solution ‍(Nx, yN , zN) ‍ such that
 
‍max(Nx, yN , zN) =u , min(Nx, yN , zN) =v ‍. To appreciate this point, consider a model
with only two decision variables ‍x ‍ and ‍y ‍, the feasible set is the triangle in
Figure 16.24, and we want to maximize ‍max(x, y)  min(x, y)‍. In plain words, we
want to find a point that is as far away from the line ‍y = x ‍ as possible. We can
see that the optimal solution is the dot in the figure, and the optimal objective
function value is very small (somewhere between 0.1 and 0.2). If we use the
aforementioned wrong approach, we have ‍u = 1, v = 0 ‍ and the optimal objec-
tive function value is 1.
A detailed explanation of the solution to the above example:
There are six possible cases: ‍x  y  z ‍, ‍x  z  y ‍, ‍y  x  z ‍, ‍y  z  x ‍,
‍z  x  y ‍, and ‍z  y  x ‍.
We solve Model 1: maximize ‍x  z ‍ subject to all the given constraints and
(1) (1) (1)
‍x  y, y  z ‍, and its optimal solution is ‍(x , y , z )‍.
274  Machine learning and data analytics for maritime studies

We solve Model 2: maximize ‍x  y ‍ subject to all the given constraints and

‍x  z, z  y ,‍ and its optimal solution is ‍(x(2) , y(2) , z(2) ).‍
We solve Model 3: maximize ‍y  z ‍ subject to all the given constraints and
‍ y  x, x  z ,‍ and its optimal solution is ‍(x(3) , y(3) , z(3) ).‍
We solve Model 4: maximize ‍y  x ‍ subject to all the given constraints and
‍ y  z, z  x ,‍ and its optimal solution is ‍(x(4) , y(4) , z(4) ).‍
We solve Model 5: maximize ‍z  y ‍ subject to all the given constraints and
z‍  x, x  y ,‍ and its optimal solution is ‍(x(5) , y(5) , z(5) ).‍
We solve Model 6: maximize ‍z  x ‍ subject to all the given constraints and
z‍  y, y  x ,‍ and its optimal solution is ‍(x(6) , y(6) , z(6) ).‍
  
The optimal solution is ‍(x(i ) , y(i ) , z(i ) ),‍ in which i‍ 2 f1, 2, 3, 4, 5, 6g‍ and
model i‍‍has the largest optimal objective function value among these six models.

References

[1] Wang S., Meng Q., Sun Z. ‘Container routing in liner shipping’. Transportation
Research Part E. 2013;49(1):1–7.
[2] Wang S. ‘A novel hybrid-­link-­based container routing model’. Transportation
Research Part E. 2014;61:165–75.
[3] Wang S., Meng Q. ‘Sailing speed optimization for container ships in a liner
shipping network’. Transportation Research Part E. 2012;48(3):701–14.
 Chapter 17
Integer optimization

In linear optimization, we assume that all of the decision variables are continu-
ous. However, in reality, some are not. For example, the number of crude oil
tankers used to transport crude oil from the Middle East to the US is an inte-
ger, and rounding down 6.5 ships to 6 ships can lead to considerable errors.
Therefore, a natural extension to linear optimization models is integer linear
optimization models, which are the same as linear optimization models except
that the decision variables can only take integer values. We also have mixed-­
integer linear optimization models, in which some decision variables can only
take integer values, and the others are continuous. We often use “integer opti-
mization models” to refer to integer linear optimization models or both inte-
ger linear optimization models and mixed-­integer linear optimization models.
However, one should keep in mind that integer linear optimization models are
not linear; in other words, integer linear optimization models belong to the cat-
egory of nonlinear optimization models.
We use Z ‍ +‍to represent the set of non-­negative integers*. Hence, ‍x 2 Z+‍means
that ‍x ‍is non-­negative and can only take integer values.

17.1 Formulation I: natural integer decision variables

Example 17.1: Suppose that there is 1 million tons of coal to transport from
Indonesia to Japan. The following bulk carriers are available:

Cost per trip

Ship type Capacity (1000 tons) Number of ships (million $)
Capesize 100 8 1
Panamax 60 8 0.8
Handymax 40 7 0.6
Handysize 10 5 0.3

* 
One often uses Z++ to represent the set of positive integers.
276  Machine learning and data analytics for maritime studies

Each bulk carrier can only complete one trip from Indonesia to Japan due to the
deadline for delivering the coal. The company needs to determine how many ships
in each type to use. Formulate an integer optimization model that minimizes the total
cost of transporting the coal.
Solution: Let ‍x ‍, ‍y , ‍‍ z ‍, and ‍w ‍be the numbers of Capesize, Panamax, Handymax,
and Handysize ships to use, respectively. The model is:

‍ min 1x + 0.8y + 0.6z + 0.3w ‍

subject to

100x + 60y + 40z + 10w  1000

x8
y8
z7
w5
x 2 Z+
y 2 Z+
z 2 Z+
‍ w 2 Z+ . ‍

Example 17.2: Suppose that there is 2 million tons of crude oil to transport from
Saudi Arabia to the US. The following crude oil tankers are available (capacity in
terms of 1000 tons and cost per trip in terms of million $):

Ship type Capacity Number Cost per trip

Very large crude carrier 200 8 2
(VLCC)
Suezmax crude tanker 120 8 1.5
Aframax crude tanker 80 7 1.2
Panamax crude tanker 60 5 1

Each ship can only complete one trip from Saudi Arabia to the US due to the
deadline for delivering the crude oil. The company needs to determine how many
ships in each type to use. Formulate an integer optimization model that minimizes
the total cost of transporting the crude oil.
Solution: Let ‍x ‍, ‍y , ‍‍ z ‍, and ‍w ‍be the numbers of VLCC, Suezmax, Aframax, and
Panamax ships to use, respectively. The model is

‍ min 2x + 1.5y + 1.2z + 1w ‍

subject to
 Integer optimization  277

200x + 120y + 80z + 60w  2000

x8
y8
z7
w5
x 2 Z+
y 2 Z+
z 2 Z+
‍ w 2 Z+ . ‍
Example 17.3: Reconsider Example 17.2. Suppose that there is a further require-
ment that at most 10 ships can be used. Formulate an integer optimization model that
minimizes the total cost of transporting the crude oil.
Solution: We only need to add the constraint ‍x + y + z + w  10‍ to the above
model.
Example 17.4: This is a ship repositioning problem. Suppose that there are
three liner services (r1, r2, and r3) on which container ships are to be deployed. The
service r1 needs four ships, r2 needs five ships, and r3 needs six ships. The available
ships are at five locations: location i‍ ‍has i‍ + 1‍ships, i‍ = 1, 2, 3, 4, 5.‍ The cost ($1000/
ship) of repositioning a ship from its current location to a service is shown below:

Location r1 r2 r3
1 11 12 13
2 25 28 37
3 0 50 40
4 20 20 0
5 31 14 2

Formulate an integer optimization model to determine how to reposition the

ships to minimize the total repositioning cost.
Solution: Let ‍xij , i = 1, 2, 3, 4, 5, j = 1, 2, 3,‍ be the number of ships to reposition
from location i‍ ‍to service ‍j .‍ The model is
min 11x11 + 12x12 + 13x13 + 25x21 + 28x22 + 27x23 + 50x32
‍ +40x33 + 20x41 + 20x42 + 31x51 + 14x52 + 2x53‍
subject to
x11 + x12 + x13  2
x21 + x22 + x23  3
x31 + x32 + x33  4
x41 + x42 + x43  5
‍ x51 + x52 + x53  6‍
278  Machine learning and data analytics for maritime studies

x11 + x21 + x31 + x41 + x51 = 4

x12 + x22 + x32 + x42 + x52 = 5
x13 + x23 + x33 + x43 + x53 = 6

‍ xij 2 Z+ , i = 1, 2, 3, 4, 5, j = 1, 2, 3.‍ 

17.2 Formulation II: 0–1 decision variables

Some integer optimization models have only binary decision variables, i.e., decision
variables which can only take the value 0 or 1. Such models are called 0–1 integer
optimization models or binary integer optimization models. 0–1 integer optimiza-
tion models have wide applications for expressing logical constraints.
Example 17.5: Consider a teaching venue allocation problem. Suppose that
there are six courses (C1–C6) to be taught in three classrooms (R1–R3). Each class-
room can only be used to teach two courses due to reservations, e.g., meetings. Not
all classrooms are suitable for all courses because of available teaching equipment,
as shown in the table below where “Y” means the classroom is suitable for the
course and “N” means not suitable. Formulate an integer optimization model to find
a feasible classroom allocation plan for all the courses.

R1 R2 R3
C1 Y Y N
C2 Y N N
C3 N N Y
C4 Y Y Y
C5 N N Y
C6 Y Y N

Solution: Let ‍xij , i = 1, 2, 3, 4, 5, 6, j = 1, 2, 3,‍ be a binary decision variable which

equals 1 if course i‍ ‍is taught in classroom ‍j ‍, and 0 otherwise. The model is

‍ min 0‍ 

subject to
x11 + x12 + x13 = 1
x21 + x22 + x23 = 1
x31 + x32 + x33 = 1
x41 + x42 + x43 = 1
x51 + x52 + x53 = 1
‍ x61 + x62 + x63 = 1‍
 Integer optimization  279

x11 + x21 + x31 + x41 + x51 + x61  2

x12 + x22 + x32 + x42 + x52 + x62  2
x13 + x23 + x33 + x43 + x53 + x63  2
x13 = x22 = x23 = x31 = x32 = x51 = x52 = x63 = 0

‍ xij 2 f0, 1g, i = 1, 2, 3, 4, 5, 6, j = 1, 2, 3.‍

Example 17.6: Consider a service with the port rotation below:
‍ Shanghai(S, 1) ! Pusan(P, 2) ! LosAngeles(L, 3) ! Shanghai(S, 1)‍

The container shipment demand is: S to P ‍qSP = 800‍ and S to L ‍qSL = 6000‍. The
revenue of transporting one TEU (twenty-­foot equivalent unit) from S to P is $200
and from S to L is $1 300. Suppose that the company needs to determine which of
the following three types of container ship to use:

Ship type Capacity (TEUs) Cost (thousand $/week)

Post-­Panamax 8 000 500
Panamax type 2 5 100 350
Panamax type 1 4 800 330

Exactly one type of ship will be used. The company aims to maximize its profit.
Formulate a mixed-­integer linear optimization model to help the company make the
decision of which type of ship to use.
Solution: Let ‍x1 , x2 , x3‍ be binary decision variables that equal 1 if and only
post-­Panamax, Panamax type 2, and Panamax type 1 ships are used, respectively,
and 0 otherwise. Let ‍ySP ‍and ‍ySL‍be the decision variables representing volumes of
containers transported from S to P and S to L, respectively. The model is
‍ max 200ySP + 1300ySL  (500, 000x1 + 350, 000x2 + 330, 000x3 )‍ 
subject to
ySP + ySL  8000x1 + 5100x2 + 4800x3
ySL  8000x1 + 5100x2 + 4800x3
ySP  800
‍ y  6000‍
SL

x1 + x2 + x3 = 1
x1 2 f0, 1g, x2 2 f0, 1g, x3 2 f0, 1g
‍ ySP  0, ySL  0.‍ 

The above example explains why we assume continuous numbers of containers in

some problems: our purpose is to decide the type of ship to use rather than to trans-
port containers according to the calculated numbers. After the ships are deployed
and the service is operated, we will know the exact demand for each week (which
will fluctuate with time and be different from the forecast value in our example) and
280  Machine learning and data analytics for maritime studies

then plan the delivery of containers by formulating integer numbers of containers

(moreover, we should also consider the type of containers, the storage of containers
on the ship, etc.).
Example 17.7: Figure 17.1 shows a liner container shipping network. The
circles represent ports. There are three candidate routes r1, r2, and r3. The capac-
ity of a ship deployed on route r1 is E1 (TEUs). E2 and E3 have similar meanings.
The volumes of containers between different OD (origin-­destination) pairs are
shown in the figure. For example, ‍q12 ‍ is the demand (TEUs/week) from port 1
to port 2. The revenue for transporting one TEU from port 1 to port 2 is ‍g12 ‍ ($/
TEU). ‍g13‍ and ‍g32 ‍ have similar meanings. Assume that the container handling
costs are 0. Suppose that the company needs to determine which ones of the
three routes r1, r2, and r3 should be operated. The costs for operating r1, r2, and
r3 are known and represented by c1, c2, and c3 ($/week), respectively. If some
routes are not profitable, then they will not be operated. Formulate a mixed-­
integer linear optimization model to help the company make the decision of
which routes to operate.
Solution: The set of OD pairs are ‍W = f(1, 2), (1, 3), (3, 2)g‍. Let ‍ Hod ‍be the set
of itineraries (paths) for OD pair ‍(o, d) 2 W ‍. To simplify the notation, we use,‍ i >‍to
represent leg i‍ ‍of route ‍r ‍. Then ‍ H12‍consists of:

h1 :< 1, 1 >
‍ h2 :< 3, 1 >!< 2, 2 >‍ 

‍ H13 consists of :‍

h3 :< 3, 1 >
‍ h4 :< 1, 1 >!< 2, 1 >‍

Figure 17.1  A container liner shipping network with direct delivery and
transshipment
 Integer optimization  281

‍ H32 consists of :‍ 
h5 :< 2, 2 >
‍ h6 :< 3, 2 >!< 1, 1 > .‍

Define ‍H := [(o,d)2W Hod ‍. Let xr be a binary decision variable that equals 1 if and
only if route ‍r ‍is operated, ‍r = 1, 2, 3‍. Let yh be the decision variable representing the
P ‍hod2 H
flow on itinerary P‍(TEUs/week). The model is
max g yh  (c1 x1 + c2 x2 + c3 x3 )
‍ (o,d)2W h2Hod ‍

subject to
P
yh  qod , (o, d) 2 W
h2Hod

y1 + y4 + y6  E1 x1 (capacity on leg < 1, 1 >)

y4  E2 x2 (capacity on leg < 2, 1 >)
y2 + y5  E2 x2 (capacity on leg < 2, 2 >)
y2 + y3  E3 x3 (capacity on leg < 3, 1 >)
y6  E3 x3 (capacity on leg < 3, 2 >)
x1 2 f0, 1g, x2 2 f0, 1g, x3 2 f0, 1g
‍ yh  0, h 2 H.‍

17.3 Formulation III: complex logical constraints

In an optimization model, the relation between the constraints is “AND,” which

means all of them must be satisfied. Binary variables can be used to formulate more
complex logic as linear constraints.
Suppose that we have three candidate constraints ‍x + y  2,‍ ‍2x  y  3,‍
3x
‍ + 4y  4‍ that do not necessarily hold. If we require that at least one of them
must hold, then we can define ‍ M ‍as a large positive number, e.g., ‍M = 1010‍ (note
that ‍M ‍is a constant rather than a decision variable), and define three binary decision
variables ‍z1 , z2 , z3‍, and the requirement can be formulated as
x + y  2  Mz1
2x  y  3  Mz2
3x + 4y  4  Mz3
z1 + z2 + z3  2
‍ z1 2 f0, 1g, z2 2 f0, 1g, z3 2 f0, 1g.‍ 

Because of ‍M ‍, we often say we use “the big-­M method” to formulate the constraints.
It might be attempting to set ‍ M ‍at a very large value to be on the safe side.
However, a larger ‍ M ‍will make the model more time-­consuming to solve. In reality,
282  Machine learning and data analytics for maritime studies

we should try to find the smallest possible ‍ M ‍. For example, if the above model
further has the following constraints: ‍0  x  2‍and 1 ‍  y  3,‍ then we know that
maxfx
‍ + y  2, 2x  y  3, 3x + 4y  4g = 8,
‍ and we can hence set ‍M = 8‍.
Example 17.8: We have a model. One of its decision variables ‍x ‍can either be
greater than or equal to 2, or less than or equal to 1. How to use the big-­M method to
formulate this requirement as linear constraints with binary variables?
Solution: The requirement is at least one of the following two constraints must
hold: ‍2  x  0‍and ‍x  1  0‍. Hence, it can be formulated as
2  x  Mz1
x  1  Mz2
z1 + z2  1
‍ z1 2 f0, 1g, z2 2 f0, 1g.‍
Example 17.9: We have a model. One of its decision variables ‍x ‍ can either be
greater than or equal to 2, or between 0 and 1 (inclusive of 0 and 1). How to use
the big-­M method to formulate this requirement as linear constraints with binary
variables?
Solution: The requirement is: ‍x  0‍ and at least one of the following two con-
straints must hold: ‍2  x  0‍and ‍x  1  0‍. Hence, it can be formulated as
2  x  Mz1
x  1  Mz2
z1 + z2  1
z1 2 f0, 1g, z2 2 f0, 1g
‍ x  0.‍

17.4 Solving mixed-integer optimization models

The model obtained by removing the integrality constraints of the variables in an
integer optimization model is called the linear programming relaxation of the integer
optimization model, which is a linear optimization model.
Example 17.10: Write down the linear programming relaxation of the model
in Example 17.1.
Solution:
‍ min 1x + 0.8y + 0.6z + 0.3w ‍ 
subject to

100x + 60y + 40z + 10w  1000

x8
y8
‍ z7
w5
x0
y0
z0
w  0.
 100x + 60y + 40z + 10w  1000
x8
y8
Integer optimization  283
z7
w5
x0
y0
z0
w  0.‍
Note that the last four non-­negativity constraints are the corresponding linear pro-
gramming relaxation of the constraints on the decision variables that they must take
non-­negative integer values.
It is easy to understand that the optimal objective function value of the linear pro-
gramming relaxation is at least as good as that of the corresponding integer or mixed-­
integer optimization model. It is even possible that the linear programming relaxation
is unbounded, while the corresponding integer optimization model is infeasible. For
example, the following mixed-­integer optimization model has no feasible solution,
while its linear programming relaxation is unbounded: maximizing ‍x + y‍ subject to
‍x  0,‍ ‍0.5  y  0.9‍and ‍y 2 Z+.‍
Most algorithms for solving integer optimization models are based on repeat-
edly solving linear programming relaxations that are related to the integer optimiza-
tion models, such as the cutting-­plane method, the branch-­and-­bound method, and
the branch-­and-­cut method. Most linear optimization solvers, such as Excel, lpsolve,
MATLAB®, and CPLEX, can also solve integer optimization models. We must bear
in mind that while linear optimization models are generally easy to solve, integer
optimization models are generally difficult to solve. It is not uncommon that a linear
optimization model can be solved in 1 s, and after imposing that all the decision
variables are integers, the model cannot be solved in 1 week. As a rule of thumb, one
might expect a model with 20 integer variables and 1 000 constraints to be solved
in 1 min and expect that a model with 100 integer variables and 1 000 constraints
cannot be solved in 1 day.

17.5 Formulation IV: challenging problems

In contrast to the fact that linear optimization models are generally easy to formu-
late, (i) it may be difficult to formulate an integer optimization model, (ii) one may
often formulate an incorrect integer optimization model, and (iii) it is possible to
formulate a bad integer optimization model when there are good formulations that
can be solved efficiently.
Example 17.11: There is a model with the following constraints
x1 = x2
x1 = x3
x1 = x4
‍ x1 2 f0, 1g, x2 2 f0, 1g, x3 2 f0, 1g, x4 2 f0, 1g.‍
Formulate one constraint to replace the first three constraints.
284  Machine learning and data analytics for maritime studies

Solution: We can use ‍3x1 = x2 + x3 + x4.‍

Note that the above example simply aims to show that there may be different
integer optimization formulations for the same problem. It does not mean that fewer
constraints are better.
Example 17.12: Reconsider Example 17.11. Which constraints can be dropped
from the set of constraints?
Solution: We can drop ‍x2 2 f0, 1g, x3 2 f0, 1g, x4 2 f0, 1g.‍
Example 17.13: This is the traveling salesman problem. Suppose that a deliv-
ery man is at node 1 (a KFC store), needs to visit ‍n  1‍ customers, denoted by
nodes ‍2, 3,    , n,‍ and returns to node 1. Define the set of nodes ‍N := f1, 2,    , ng‍
and the set of arcs ‍A := f(i, j), i 2 N, j 2 N, i ¤ jg.‍ The distance of arc ‍(i, j)‍ is ‍dij .‍
Develop an integer optimization model to determine the shortest tour to visit all
customers.
Beginners might work as follows: let ‍xij ‍be a binary variable which equals 1 if
and only if arc ‍(i, j)‍is traversed, and the model is
P
min dij xij
‍ (i,j)2A ‍
subject to
P
xij = 1, i 2 N
j2N\fig
P
xij = 1, j 2 N
i2N\fjg

‍ xij 2 f0, 1g, (i, j) 2 A. ‍

The above formulation is incorrect because its solution may contain subtours, see
the example in Figure 17.2.
Solution: Let ‍xij ‍be a binary variable which equals 1 if and only if arc ‍(i, j)‍is
traversed, and let ui represent the number of customers that have been served after
visiting i‍ ‍. The purpose of the decision variables ui is to eliminate subtours. The
model is
P
min dij xij
‍ (i,j)2A ‍

Figure 17.2   Subtours in TSP

 Integer optimization  285

subject to
P
xij = 1, i 2 N
j2N\fig
P
xij = 1, j 2 N
i2N\fjg

xij 2 f0, 1g, (i, j) 2 A

uj  ui + 1  (n  1)(1  xij ), i 2 N, j 2 N \ f1g, j ¤ i
ui  1, i 2 N \ f1g
‍ u1 = 0. ‍

Example 17.14: In the traveling salesmanP problem, write down the model for the
case of ‍n = 4‍. You are not allowed to use “‍ ‍” or “‍8‍” except in the non-­negativity/
binary constraints.
Solution: Let ‍xij ‍ be a binary variable which equals 1 if and only if arc ‍(i, j)‍ is
traversed, and let ui represent the number of customers that have been served after
visiting i‍ ‍. The model is
min d12 x12 + d13 x13 + d14 x14 + d21 x21 + d23 x23 + d24 x24 +
‍ d31 x31 + d32 x32 + d34 x34 + d41 x41 + d42 x42 + d43 x43 ‍

subject to
x12 + x13 + x14 = 1
x21 + x23 + x24 = 1
x31 + x32 + x34 = 1
x41 + x42 + x43 = 1
x21 + x31 + x41 = 1
x12 + x32 + x42 = 1
x13 + x23 + x43 = 1
x14 + x24 + x34 = 1

‍ xij 2 f0, 1g, (i, j) 2 A‍

u2  u1 + 1  (n  1)(1  x12 )
u2  u3 + 1  (n  1)(1  x32 )
u2  u4 + 1  (n  1)(1  x42 )
u3  u1 + 1  (n  1)(1  x13 )
u3  u2 + 1  (n  1)(1  x23 )
u3  u4 + 1  (n  1)(1  x43 )
u4  u1 + 1  (n  1)(1  x14 )
u4  u2 + 1  (n  1)(1  x24 )
‍ u4  u3 + 1  (n  1)(1  x34 )
u2  1
u3  1
u4  1
u1 = 0.
 u3  u1 + 1  (n  1)(1  x13 )
u3  u2 + 1  (n  1)(1  x23 )
u3  u4 + 1  (n  1)(1  x43 )
u4  u1 + 1  (n  1)(1  x14 )
u4  u2 + 1  (n  1)(1  x24 )
286  uMachine learning and data analytics for maritime studies
4  u3 + 1  (n  1)(1  x34 )

u2  1
u3  1
u4  1
u1 = 0.‍
Example 17.15: Revisit the shortest path problem. A transportation network has
a set of nodes ‍N := f1, 2,    , ng‍ and a set of arcs ‍A := f(i, j), i 2 N, j 2 N, i ¤ jg‍.
The distance of arc ‍(i, j)‍ is ‍dij ‍. Develop an integer optimization model to deter-
mine the shortest path from node 1 to node ‍n‍. Note the differences between the
shortest path problem and the TSP ( travelling salesman problem); in the short-
est path problem, not all nodes have to be visited, and the person does not return
to node 1.
Solution: Let ‍xij ‍ be a binary variable which equals 1 if and only if arc ‍(i, j)‍ is
traversed. The model is
P
min dij xij
‍ (i,j)2A ‍
subject to
P
x1j = 1
j2N\f1g
P
xin = 1
i2N\fng
P P
xij = xji , i 2 N \ f1, ng
j2N\fig j2N\fig

‍ xij 2 f0, 1g, (i, j) 2 A.‍

Note that the second constraint can be removed as it is implied by other constraints.
Note further that you can add the following constraints to the model if you like:
‍xi1 = 0, i 2 N \ f1g‍and ‍xnj = 0, j 2 N \ fng.‍
Example 17.16: In the shortest path problem,P write down the model for the
case of ‍n = 4‍. You are not allowed to use “‍ ”‍ or “‍8‍” except in the non-­negativity/
binary constraints.
Solution: The set of arcs are ‍A := f(1, 2), (1, 3), (1, 4), (2, 3), (2, 4), (3, 2), (3, 4)g.‍
Let ‍xij ‍be a binary variable which equals 1 if and only if arc ‍(i, j) 2 A ‍is traversed. The
model is

‍ min d12 x12 + d13 x13 + d14 x14 + d23 x23 + d24 x24 + d32 x32 + d34 x34‍
subject to
x12 + x13 + x14 = 1
x14 + x24 + x34 = 1
x23 + x24 = x12 + x32
x32 + x34 = x13 + x23
‍ xij 2 f0, 1g, (i, j) 2 A.‍ 
 Integer optimization  287

Example 17.17: Ships need to be deployed on three services r1, r2, and r3 to main-
tain a weekly frequency. The operating cost of a service given the number of ships
deployed is shown below.

Cost (million $/week) Five ships Six ships Seven ships

r1 1.5 1.3 1.1
r2 2.0 1.9 1.6
r3 2.3 2.2 2.0

Suppose that the company has ten ships and can charter in additional ships at
the cost of 0.1 million $/week per ship. Develop a model to determine how to deploy
ships on the three services at minimum cost.
Solution: Let xr and yr be the numbers of owned ships and chartered ships to
deploy on route ‍r = 1, 2, 3,‍ respectively. Let ‍zrs‍be a binary variable that equals 1 if
and only if ‍s = 5, 6, 7‍ships are deployed on route ‍r = 1, 2, 3.‍ The model is

min
00.1(y1 + y2 + y3 )+
1.5z15 + 1.3z16 + 1.1z17 + 2z25 + 1.9z26 + 1.6z27 + 2.3z35 + 2.2z36 + 2z37
‍ ‍
subject to
P
7
xr + yr = szrs , r = 1, 2, 3
s=5
P
7
zrs = 1, r = 1, 2, 3
s=5

xr 2 Z+ , r = 1, 2, 3
P
3
xr  10
r=1

yr 2 Z+ , r = 1, 2, 3
‍ zrs 2 f0, 1g, r = 1, 2, 3, s = 5, 6, 7.‍ 

The problem with the above formulation is that it has a large number of symmetrical
optimal solutions. For example, if ‍(x1 , y1 , x2 , y2 , x3 , y3 ) = (4, 1, 4, 2, 2, 5)‍is an opti-
mal solution, then ‍(x1 , y1 , x2 , y2 , x3 , y3 ) = (3, 2, 4, 2, 3, 4)‍ is also an optimal solu-
tion. In fact, there will be 35 optimal solutions. Generally speaking, having so many
symmetrical feasible and optimal solutions will make the model very difficult to
solve (the reason is beyond the scope of the subject) Readers can refer to Reference
[1] to see a comparison of efficiency of the two models. A better formulation is as
follows.
Solution: Let mr be the number of ships (including owned and chartered ships)
to deploy on route ‍r = 1, 2, 3‍. Let ‍x ‍be the total number of owned ships deployed on
the three routes and ‍y ‍be the total number of chartered ships deployed on the three
288  Machine learning and data analytics for maritime studies

routes. Let ‍zrs‍ be a binary variable that equals 1 if and only if ‍s = 5, 6, 7‍ ships are
deployed on route ‍r = 1, 2, 3.‍ The model is

min
00.1y+
‍ 1.5z15 + 1.3z16 + 1.1z17 + 2z25 + 1.9z26 + 1.6z27 + 2.3z35 + 2.2z36 + 2z37‍

subject to
P
7
mr = szrs , r = 1, 2, 3
s=5
P7
zrs = 1, r = 1, 2, 3
s=5
P3
mr = x + y
r=1

x  10
mr 2 Z+ , r = 1, 2, 3
x 2 Z+
y 2 Z+
‍ zrs 2 f0, 1g, r = 1, 2, 3, s = 5, 6, 7.‍ 
We can further prove that the integrality constraints on ‍x ‍and ‍y ‍can be removed.
Moreover, for some problem, natural integer optimization formulations have
an exponential number of variables, while more clever formulations have only a
polynomial number of variables.

17.6 Formulation V: linearizing binary variables multiplied by

another variable

Given ‍z1 2 f0, 1g, z2 2 f0, 1g,‍ the constraint ‍z  z1 z2‍ is equivalent to
‍z  z1 + z2  1, z  0‍; the constraint ‍z  z1 z2‍ is equivalent to ‍z  z1 , z  z2‍; and
the constraint ‍z = z1 z2‍is equivalent to ‍z  z1 , z  z2 , z  z1 + z2  1, z  0‍.
Given ‍z 2 f0, 1g, 0  x  M ,‍ the constraint ‍y  zx ‍ is equivalent to
‍y  0, y  x  M(1  z);‍ the constraint ‍y  zx ‍ is equivalent to ‍y  x, y  Mz ;‍ and
the constraint ‍y = zx ‍is equivalent to ‍y  0, y  x  M(1  z), y  x, y  Mz .‍
Example 17.18: A model has decision variables ‍z1 2 f0, 1g, z2 2 f0, 1g, 0  x  M,‍
an objective function that minimizes ‍3z1 z2  4z1 x‍, and some linear constraints. How do
you linearize the objective function?
Solution: We introduce two new variables u1 and u2. The model is
‍ min 3u1  4u2‍ 
 Integer optimization  289

subject to
u 1 = z1 z2 ,
‍ u2 = z1 x,‍ 
and other relevant constraints. This model is equivalent to:
‍ min 3u1  4u2‍ 
subject to
u1  z1 z2 ,
‍ u2  z1 x,‍ 
and other relevant constraints. This model is equivalent to:
‍ min 3u1  4u2‍
subject to
u1  z1 + z2  1,
u1  0,
u2  x,
‍ u2  Mz1 ,‍ 
and other relevant constraints.

17.7 Practice

Example 17.19: A student has at most 20 days for preparing the final exams of four
subjects. Each subject is worth six credit points. The student needs to determine how
many days to spend on each subject. It is required that the number of days allocated
to each subject must be an integer.
Suppose that the final score of a subject is proportional to the number of
days spent on it. Let ci be the scores the student can get if he spends 1 day on
subject i‍ = 1, 2, 3, 4.‍ For instance, if ‍c1 = 11‍, then the student can get 99 marks if
he spends 9 days on subject 1; as the student cannot get more than 100 marks,
the student will not spend more than 9 days on the subject. We therefore let
T‍ i ‍ be the maximum days the student spends on subject i‍ = 1, 2, 3, 4‍ is the larg-
est integer that does not exceed 100 . The student knows the values of ci and T
‍ i ,‍
‍ ci ‍
i‍ = 1, 2, 3, 4 .‍
The student wants to maximize his average mark. How should he allocate time
to the four subjects?
Solution: Let ‍xi , i = 1, 2, 3, 4,‍ be the number of days spent on subject i‍ ‍. The
model is
c1 x1 + c2 x2 + c3 x3 + c4 x4
max
‍ 4 ‍
290  Machine learning and data analytics for maritime studies

subject to

x1 + x2 + x3 + x4  20
xi  Ti , i = 1, 2, 3, 4
‍ xi 2 Z+ , i = 1, 2, 3, 4.‍ 

Example 17.20: In the context of Example 17.19, the student must make sure that
he passes all the subjects (at least 50 marks). How should he allocate time to the four
subjects in order to maximize his average mark?
Solution: Let ‍xi , i = 1, 2, 3, 4‍, be the number of days spent on subject i‍ ‍. The
model is
c1 x1 + c2 x2 + c3 x3 + c4 x4
max
‍ 4 ‍
subject to
x1 + x2 + x3 + x4  20
ci xi  50, i = 1, 2, 3, 4
xi  Ti , i = 1, 2, 3, 4
‍ xi 2 Z+ , i = 1, 2, 3, 4.‍

Example 17.21: In the context of Example 17.19, the student must make sure
that he passes all the subjects. How should he allocate time to the four sub-
jects in order to get the most number of “HD” (at least 85 marks)? (Hint: Let
‍xi , i = 1, 2, 3, 4,‍ be the number of days spent on subject i‍ ‍. Let ‍zi , i = 1, 2, 3, 4,‍ be
a binary variable that equals 1 if and only if he gets HD for subject i‍ ‍. Then we
maximize ‍z1 + z2 + z3 + z4 .‍ We may have the constraint ‍ci xi  85zi ,‍ so that zi can
be 1 only if ‍ci xi  85.‍)
Solution: Let ‍xi , i = 1, 2, 3, 4‍, be the number of days spent on subject i‍ ‍. Let
‍zi , i = 1, 2, 3, 4‍, be a binary variable that equals 1 if and only if he gets HD for sub-
ject i‍ ‍. The model is

‍ max z1 + z2 + z3 + z4‍
subject to
x1 + x2 + x3 + x4  20
ci xi  50, i = 1, 2, 3, 4
ci xi  85zi , i = 1, 2, 3, 4
xi  Ti , i = 1, 2, 3, 4
xi 2 Z+ , i = 1, 2, 3, 4
‍ zi 2 f0, 1g, i = 1, 2, 3, 4.‍
Example 17.22: In the context of Example 17.19, develop a model to help the stu-
dent check whether he can get “D” (at least 75 marks) for all the subjects.
 Integer optimization  291

Solution: Let ‍xi , i = 1, 2, 3, 4,‍ be the number of days spent on subject i‍ ‍. The
model is
‍ min 0‍ 
subject to
x1 + x2 + x3 + x4  20
ci xi  75, i = 1, 2, 3, 4
xi  Ti , i = 1, 2, 3, 4
‍ xi 2 Z+ , i = 1, 2, 3, 4.‍
If the above model has a feasible solution, he can get “D” for all the subjects, oth-
erwise he cannot.
Example 17.23: In the context of Example 17.19, develop a model to help the
student check whether he can get at least two “D” and two “HD” (i.e., two “D” and
two “HD,” one “D” and three “HD,” or four “HD”).
Solution: Similar to Example 17.23, this question requires that the score for
each subject should at least be 75 and checks whether the maximum number of
“HD” exceeds two. Let ‍xi , i = 1, 2, 3, 4‍, be the number of days spent on subject i‍ ‍.
Let ‍zi , i = 1, 2, 3, 4‍, be a binary variable that equals 1 if and only if he gets HD for
subject i‍ ‍. The model is

‍ max z1 + z2 + z3 + z4‍ 
subject to
x1 + x2 + x3 + x4  20
ci xi  75, i = 1, 2, 3, 4
ci xi  85zi , i = 1, 2, 3, 4
xi  Ti , i = 1, 2, 3, 4
xi 2 Z+ , i = 1, 2, 3, 4
‍ zi 2 f0, 1g, i = 1, 2, 3, 4.‍ 
If the optimal objective function value is greater than or equal to 2, the student can
get at least two “D” and two “HD,” otherwise he cannot.
Example 17.24: In the context of Example 17.19, it may not be reasonable to
assume that the score of a subject is proportional to the number of days spent on it.
Therefore, we let ‍cij ‍ be the total score of subject i‍ = 1, 2, 3, 4‍ the student can get if
‍j = 0, 1,    , 20‍days is spent on it. Of course, ‍0  ci0  ci1  ci2      ci,20  100.‍
The student knows the values of ‍cij .‍ How should he allocate time to the four subjects in
order to maximize his average mark? (Hint: Let ‍yij , i = 1, 2, 3, 4‍and ‍j = 0, 1,    , 20‍,
be a binary variable that equals 1 if and only if ‍j ‍days is spent on subject i‍ ‍.)
Solution: Let ‍yij , i = 1, 2, 3, 4‍ and ‍j = 0, 1,    , 20‍, be a binary variable that
equals 1 if and only if ‍j ‍days is spent on subject i‍ ‍. The model is:
P4 P20
i=1 j=0 cij yij
‍ max 4 ‍ 
292  Machine learning and data analytics for maritime studies

subject to
P
4 P
20
jyij  20
i=1 j=0
P
20
yij = 1, i = 1, 2, 3, 4
j=0

‍ yij 2 f0, 1g, i = 1, 2, 3, 4, j = 0, 1,    , 20.‍

Example 17.25: ‍M ‍is a given number. We have the constraints ‍y  x1 z + x2 z2 + x3 z3 ,‍

‍z 2 f0, 1g, 0  x1  M,‍‍0  x2  M, 0  x3  M ‍. How to linearize the first one?
Solution: Since ‍ z ‍is binary, ‍ z = z2 = z3‍ . Therefore, the con-
straint is ‍y  x1 z + x2 z + x3 z = (x1 + x2 + x3 )z ‍, which is equivalent to
‍y  x1 + x2 + x3  3M(1  z),‍‍y  0‍.
Example 17.26: ‍ M ‍is a given positive number. We have the constraints
‍y = xu, u 2 Z+ , 0  x  M, u  M ‍. How to linearize the first one?
Solution: Define Pbinary P variables ‍zi 2 f0, 1g, i = 0, 1    M ‍. The constraints can
be written as ‍y = x M i=0 iz i , M
‍ ‍ i=0 zi = 1, 0  x  M .‍
Define ‍ui = xzi , i = 0, 1    M ‍, which isPequivalent to ‍ui  0, ui  x  M(1  zi ),‍
M
‍ui  x, ui  Mzi ,‍i‍ = 0, 1    M ‍. Then ‍y = i=0 iui .‍
In sum, the linearized constraints are
P
M
y= iui
i=0

ui  0, i = 0, 1    M
ui  x  M(1  zi ), i = 0, 1    M
ui  x, i = 0, 1    M
ui  Mzi , i = 0, 1    M
PM
zi = 1
i=0

zi 2 f0, 1g, i = 0, 1    M
‍ 0  x  M. ‍ 

Example 17.27: Suppose that there is a mixed-­integer linear optimization solver

that can address mixed-­integer linear optimization models with at most one inte-
ger variable in each constraint. For example, if ‍x 2 R, y 2 R, z1 2 Z , z2 2 Z ‍, then
the constraint ‍x + 2y  3z1  4‍ has only one integer variable and therefore can be
addressed; the constraint ‍x + z2 = 4‍ can also be addressed. How do you use this
solver to address a mixed-­integer linear optimization model with at most one integer
variable in each constraint except the following one ‍x + 2y  3z1 + 5z2  4‍?
Solution: Define a continuous decision variable ‍u 2 R‍. The constraint is equiva-
lent to
x + 2y  3z1 + 5u  4
‍ u = z2 .‍
 Integer optimization  293

Example 17.28: Consider a different version of the shortest path problem. A

transportation network has a set of nodes ‍N := f1, 2,    , ng‍ and a set of arcs
‍A := f(i, j), i 2 N, j 2 N, i ¤ jg,‍ ‍n  10‍. The distance of arc ‍(i, j)‍ is ‍dij .‍ We need to
find the shortest path from node 1 to node ‍n‍subject to the requirement that at least
three nodes between 1 and ‍n‍must be visited. For instance, the tour 1‍ ! 5 ! 2 ! n‍
is not valid. Develop an integer optimization model to solve the problem.
Solution: The model below is incorrect because the resulting solution may con-
tain subtours. Let ‍xij ‍ be a binary variable which equals 1 if and only if arc ‍(i, j)‍ is
traversed. Let zi be a binary variable P which equals 1 if and only if node i‍ 2 N \ f1, ng‍
is visited. The model is min dij xij
subject to ‍ (i,j)2A ‍
P
x1j = 1
j2N\f1g
P
xin = 1
i2N\fng
P
xij = zi , i 2 N \ f1, ng
j2N\fig
P
xji = zi , i 2 N \ f1, ng
j2N\fig
P
zi  3
i2N\f1,ng

xij 2 f0, 1g, (i, j) 2 A

‍ zi 2 f0, 1g, i 2 N \ f1, ng.‍
The solution below is correct.
Let zi be a binary variable which equals 1 if and only if node i‍ ‍ is visited,
i‍ 2 N \ f1, ng,‍ let ‍xij ‍be a binary variable which equals 1 if and only if arc ‍(i, j)‍is tra-
versed, and let ui be auxiliary variables for eliminating subtours, i‍ 2 N ‍. The model is
P
min dij xij
‍ (i,j)2A ‍
subject to
P
x1j = 1
j2N\f1g
P
xin = 1
i2N\fng
P
xij = zi , i 2 N \ f1, ng
j2N\fig
P
xij = zj , j 2 N \ f1, ng
i2N\fjg
P
zi  3
‍ i2N\f1,ng ‍
xij 2 f0, 1g, (i, j) 2 A
uj  ui + 1  (n  1)(1  xij ), i 2 N, j 2 N \ f1g, j ¤ i
zi 2 f0, 1g, i 2 N \ f1, ng
‍ ui  0, i 2 N. ‍
294  Machine learning and data analytics for maritime studies

Example 17.29: A student has at most 20 days for preparing the final exams of four
subjects. Each subject is worth six credit points. The student needs to determine how
many days to spend on each subject. It is required that the number of days allocated
to each subject must be an integer.
Suppose that the final mark of a subject is proportional to the number of days spent
on it. Let ci be the mark the student can get if he spends 1 day on subject i‍ = 1, 2, 3, 4‍.
For instance, if ‍c1 = 11‍, then the student can get 99 marks if he spends 9 days on sub-
ject 1; as the student cannot get more than 100 marks, the student will not spend more
than 9 days on the subject. We therefore let T ‍ i‍ be the maximum number of days the
student spends on subject i‍ = 1, 2, 3, 4‍ is the largest integer that does not exceed 100 .
‍ ci ‍
The student knows the values of ci and T ‍ i‍, i‍ = 1, 2, 3, 4‍. All ci, i‍ = 1, 2, 3, 4‍, are inte-
gers. Therefore, the final mark of a subject must also be an integer.
The student wants to maximize his average mark while (i) not failing any subject
(i.e., the final mark of each subject is at least 50) and (ii) not getting a “credit” in any
subject (i.e., the final mark of each subject is either less than or equal to 64, or greater
than or equal to 75). In sum, the final mark of each subject is either between 50 (inclu-
sive) and 64 (inclusive), or at least 75 (inclusive). Formulate an integer linear optimi-
zation model to help the student allocate time to the four subjects to achieve his goal.
Solution: Let ‍xi , i = 1, 2, 3, 4‍, be the number of days spent on subject i‍ ‍. The
model is
c1 x1 + c2 x2 + c3 x3 + c4 x4
max
‍ 4 ‍
subject to
x1 + x2 + x3 + x4  20
ci xi  50, i = 1, 2, 3, 4
ci xi  64 or ci xi  75, i = 1, 2, 3, 4
xi  Ti , i = 1, 2, 3, 4
‍ xi 2 Z+ , i = 1, 2, 3, 4.‍ 
To linearize the “or” constraints, we introduce new binary variables yi and zi,
i‍ = 1, 2, 3, 4‍. The integer linear optimization model is
c1 x1 + c2 x2 + c3 x3 + c4 x4
max
‍ 4 ‍
subject to
x1 + x2 + x3 + x4  20
ci xi  50, i = 1, 2, 3, 4
ci xi  64  100yi , i = 1, 2, 3, 4
75  ci xi  100zi , i = 1, 2, 3, 4
yi + zi  1, i = 1, 2, 3, 4
xi  Ti , i = 1, 2, 3, 4
‍ xi 2 Z+ , i = 1, 2, 3, 4
yi 2 f0, 1g, i = 1, 2, 3, 4
zi 2 f0, 1g, i = 1, 2, 3, 4.
 ci xi  50, i = 1, 2, 3, 4
ci xi  64  100yi , i = 1, 2, 3, 4
75  ci xi  100zi , i = 1, 2, 3, 4
yi + zi  1, i = 1, 2, 3, 4
xi  Ti , i = 1, 2, 3, 4
Integer optimization  295
xi 2 Z+ , i = 1, 2, 3, 4
yi 2 f0, 1g, i = 1, 2, 3, 4
zi 2 f0, 1g, i = 1, 2, 3, 4.‍ 

Example 17.30: A salesman is at node 1. There are ‍n  1‍customers, denoted by nodes

‍2, 3,    , n‍, ‍n  4‍. The salesman only needs to visit ‍n  3‍customers (the remaining
two customers can be visited by other salesmen) and then return to node 1. Define
the set of nodes ‍N := f1, 2,    , ng‍and the set of arcs ‍A := f(i, j), i 2 N, j 2 N, i ¤ jg.‍
The distance of arc ‍(i, j) 2‍is ‍dij .‍ Develop an integer optimization model to help the
salesman determine which ‍n  3‍ customers to visit and in what sequence to visit
them so that the total distance of visiting them and then returning to node 1 is the
shortest.
Solution: Let zi be a binary variable which equals 1 if and only if customer i‍ ‍is
visited, i‍ = 2, 3    , n‍, let ‍xij ‍ be a binary variable which equals 1 if and only if arc
‍(i, j)‍is traversed, and let ui represent the number of customers that have been served
after visiting i‍ ‍. The model is
P
min dij xij
‍(i,j)2A ‍

subject to
P
x1j = 1
j2N\f1g
P
xi1 = 1
i2N\f1g
P
xij = zi , i 2 N \ f1g
j2N\fig
P
xij = zj , j 2 N \ f1g
i2N\fjg

xij 2 f0, 1g, (i, j) 2 A

uj  ui + 1  (n  1)(1  xij ), i 2 N, j 2 N \ f1g, j ¤ i
ui  1, i 2 N \ f1g
u1 = 0
zi 2 f0, 1g, i 2 N \ f1g
P
n
zi = n  3.
‍ i=2 ‍

Example 17.31: (i) An optimization model has decision variables

‍z1 2 f0, 1g, z2 2 f0, 1g, z3 2 f0, 1g‍and ‍z 2 f0, 1g.‍ One of its constraints is ‍z  z1 z2 z3.‍
How do you linearize this constraint?
(ii) An optimization model has decision variables ‍z1 2 f0, 1g, z2 2 f0, 1g, z3 2 f0, 1g‍
and ‍z 2 f0, 1g.‍ One of its constraints is ‍z = z1 z2 z3.‍ How do you linearize this constraint?
(iii)Anoptimizationmodelhasdecisionvariables ‍z1 2 f0, 1g, z2 2 f0, 1g, z3 2 f0, 1g‍
and ‍z 2 f0, 1g.‍ One of its constraints is ‍z  z1 z2 z3.‍ How do you linearize this constraint?
296  Machine learning and data analytics for maritime studies

Solution: (i) ‍z  z1 + z2 + z3  2‍.

(ii) ‍z  z1 + z2 + z3  2, z  z1 , z  z2 , z  z3‍.
(iii) ‍z  z1 , z  z2 , z  z3‍.

Reference

[1] Wang S., Wang T., Meng Q. ‘A note on liner ship fleet deployment’. Flexible
Services and Manufacturing Journal. 2011;23(4):422–30.
 Chapter 18
Conclusion

18.1 Summary of this book

Machine Learning and Data Analytics in Maritime Studies: Models, Algorithms,

and Applications aims to explore the fundamental principle of analyzing practical
problems in maritime transportation using data-­driven models, especially using
machine learning (ML) models and operations research models. The book first
introduces the state-­ of-­
the-­
art data-­
enabled methodologies, technologies, and
applications in maritime transportation in a way that is easy to understand by mar-
itime researchers and practitioners. To achieve this, plain words are used to pre-
sent the algorithms and models, and real examples of ship inspection by port state
control and operations planning of container ships are accompanied. By doing so,
readers are expected to learn how to solve practical problems in maritime trans-
portation by data-­driven models while taking shipping domain knowledge and
black box model explanations into consideration.
A total of 5 parts comprising 17 chapters are contained in this book. In Part
I, the status quo of maritime transportation, including key players in the shipping
industry, in addition to an overview of container liner shipping, is introduced in
Chapter 1. In Chapter 2, the comprehensive issues in ship inspection by port state
control are covered after discussing key issues in maritime transport, and the data-
set used in this book is also introduced.
Part II aims to give an overview of data-­driven models. In Chapter 3, typi-
cal predictive problems in maritime transport are first discussed, and popular
data-­driven approaches to solving them are also covered. In Chapter 4, detailed
procedure of developing ML models to address maritime transport problems is
introduced, where comprehensive explanations and illustrations to each step are
also given.
Part III introduces popular ML models that are used to tackle practical
problems in maritime transport. Especially, linear regression models (Chapter
5), Bayesian networks (Chapter 6), support vector machine (Chapter 7), artifi-
cial neural network (Chapter 8), tree-­based models (Chapter 9), association rule
learning (Chapter 10), and cluster analysis (Chapter 11) are introduced in details.
Applications and result explanations of all the above models to predict ship risk in
port state control are given as examples, and the core code for model construction
is also given.
298  Machine learning and data analytics for maritime studies

Part IV explores data-­driven models to address practical issues in maritime

transportation more deeply. Chapter 12 first summarizes topics in maritime trans-
portation research and popular research methods to solve practical problems.
Issues of adopting data-­driven models to address these practical problems are also
discussed from the perspectives of data, model, user, and target. Chapter 13 fur-
ther discusses the issue of incorporating shipping domain knowledge into data-­
driven model developing using examples in port state control and ship energy
efficiency prediction. Chapter 14 further illustrates issues regarding explanations
of ML models within the context of maritime transportation, where examples of
decision tree explanation and using SHAP to explain ML models in black box
nature are given as examples.
Part V introduces another important data analytics approach based on opera-
tions research, which is also widely used to address practical problems in maritime
transportation. Chapter 15 introduces linear optimization techniques as the foun-
dation of operations research. Chapter 16 presents more advanced linear optimiza-
tion techniques. Chapter 17 covers integer optimization.

18.2 Future research agenda

Several future research directions are discussed in this section. It should be noted
that these topics are just the tip of the iceberg. There is still a long way to go to
realize digitization transformation in the traditional maritime industry.

1. Data play the key role and act as the foundation of constructing data-­driven
models to solve practical problems in maritime transport. However, there are
digital inefficiencies in the maritime transport chains, where there are rare pub-
lic data sources while data provided by commercial companies can be very
expensive. Efforts should be put to promote digital collaborations between ship
to port, port to port, and ship to hinterland for data sharing. Data-­driven models
developed to address practical problems in maritime transport should also com-
ply with shipping domain knowledge and are expected to take such knowledge
explicitly into the procedure of data construction.
2. Data-­driven models are usually of black box nature, which means that they are
opaque to model users or even model developers regarding how the predic-
tion model works and why a certain prediction is made. This issue should also
be taken into account when developing and explaining data-­driven prediction
models to improve model transparency and acceptance.
3. In many practical problems, accurate prediction is far from enough. Instead, it
is expected that predictions given by data-­driven models should be input to the
following optimization problems to prescribe better decisions. Unfortunately,
there is still a big gap between making a good prediction and making a good
decision, where future research efforts should be put into.
4. Up to now, as covered in Chapter 1, there are only a few maritime transporta-
tion problems that have been dealt with by ML approaches, and most of these
 Conclusion  299

problems are prediction-­based. Even if there are such prediction models, real-
izing accurate and real-­time prediction is still rare and difficult. There are sev-
eral important issues from both shipping and port sides that have not yet been
addressed, such as vessel routing and scheduling, fleet planning and develop-
ment, shipping network design, port congestion prediction and management,
port resource allocation management, and port safety and security management.
In future research, these issues are expected to be addressed.
This page intentionally left blank
 Index

Abuja MoU 11, 12 Black-box ML models

Accumulated error backpropagation advantages 195–196
100 explanation 193–198
Accuracy 44 elements 196–197
Adaboost 122–125 forms 198–200
Agglomerative algorithm 157–162 intrinsic explanation model using DT
Apriori algorithm 133–141 200–202
Area under ROC curve (AUC) 46–47 SHAP method 202–208
Artificial neural network (ANN) Black Sea MoU 11, 12
concepts 93–97 Boosting tree 125–130
model Bulk carrier 1
hyperparameters 100–101, 103
training 97–100
structure 93–97 C4.5 algorithm 109–110
Asia Pacific Computerized Information Central China 1 (CC1) service 6
System (APCIS) 171 Central China 2 (CC2) service 5, 7
Association rule learning 131 Centroid-linkage clustering 149
Apriori algorithm 133–141 Chi-squared test (c2 test) 41–42
definition 132–133 Classic clustering algorithm 151, 152
FP-growth algorithm 141–142 Classification and regression tree
Automatic identification system (AIS) (CART) 110–113
33, 167–169 Classification decision tree
C4.5, 109–110
classification and regression tree
Backpropagation algorithm 98 (CART) 110–113
Backward sequential elimination and iterative dichotomizer 3, 106–109
joining (BSEJ) 69 Classification society 4
Backward sequential elimination (BSE) Cluster analysis 145
69 clustering algorithm performance
Baltic dry index (BDI) 168 evaluation metrics for 149–151
Base learners 121 agglomerative algorithm 157–162
Bayesian network (BN) classifiers 73–77 classic algorithm 151, 152
BN classifiers 73–77 density-based spatial clustering
Naive Bayes classifier 63–68 of applications with noise
semi-naive Bayes classifiers 68–73 (DBSCAN) 154–157
Binary variables 290–291 K-means algorithm 151, 153
302  Machine learning and data analytics for maritime studies

distance measure eXtreme Gradient Boosting (XGBoost)

of clusters 148–149 182, 184–186
of examples 145–148
Complete-linkage clustering 149
Container liner shipping 5–7 Feature engineering 34
Container ship 2–3 Chi-squared test (c2 test) 41–42
Convex integration 186–190 data cleaning 35, 36
Counterfactual explanation 199 encoding 38, 40
Cross validation 47 extraction 35
Cubic law 22 mutual information (MI) 42
Pearson’s correlation coefficient
(Pearson’s r ) 41
Data cleaning 35, 36 pre-processing 37–39
Data-driven models recursive feature elimination
adoption issues 169–170 (RFE) 42
data 170–171 scaling 38, 40
model 171–172 selection 40–41
target 174–175 based on regularization 42
user 172–174 on importance 42
ML modeling (see ML modeling) Spearman’s rank coefficient
statistical modeling 23–25 (Spearman’s r) 41
theory-based modeling vs., 23 Feature extraction 35
Davies–Bouldin index (DBI) 149–150 Flag state 3–4
Decision tree (DT) 105–106 Forward sequential elimination (FSE) 69
advantages 115 Fowlkes and Mallows index (FMI)
black-box ML models 200–202 150–151
boosting and bagging based on 128 Freight forwarder 4
construction algorithms 117 Frequent pattern (FP)-growth algorithm
disadvantages 115–116 141–142
Decision variables 211, 212, 215,
280–283
Deep learning models 29, 93, 102, 103 Gaussian kernel 89, 91
Deep neural networks (DNN) 29, Gini index 110–111
100, 103 Gradient boosting decision tree
Deficiency number 18, 52–54, 61, 62, (GBDT) 125–130
64, 67, 184–186, 200–207 Grid search 49
Density-based spatial clustering
of applications with noise
(DBSCAN) algorithm 154–157 Handymax carriers 38
Dropout 100 Hard margin support vector machine
Dunn index (DI) 150 79–83

Elastic net regression 23 Integer optimization 277–298

Euclidean distance 146–147 Intrinsic explanation 199
Excel 227–229 Iterative dichotomizer 3, 106–109
 Index  303

Kernel principal components analysis 35 data collection 32–34

Kernel trick 86–89 feasibility assessment 32
k-fold cross validation 47 feature engineering (see Feature
K-means algorithm 151, 153, 158, engineering)
160–162 interpretation 50
model construction 43–48
model refinement 48–50
Lagrange multipliers 85 problem specification 32
Laplacian kernel 89 Manhattan distance 146
Lasso regression 23, 25 Maritime Labor Convention 2006
Leaky ReLU activation function 98 (MLC2006) 10
Least absolute shrinkage and selection Maritime transport
operator (LASSO) regression key issues
61–62 marine pollution control 10
Least squares method 51–52 maritime safety management 9–10
Leave-one-out cross validation 47 seafarers’ management 10
Linear discriminant analysis (LDA) 35 methods and applications 165–169
Linear kernel 86, 88, 89 overview 1, 2
Linear optimization predictive problem 21–24
advanced 252–276 research topics 163–164
basics 211–214 Mean absolute error (MAE) 43
classification 215–217 Mean absolute percentage error
dummy nodes and links 247–250 (MAPE) 43
equivalence between different Mean normalization 40
formulations 217–220 Mean-squared error (MSE) 43, 51–52, 201
Excel 227–229 Mediterranean MoU 11, 12
network flow optimization 233–247 Memorandum of Understanding (MoU)
for nonlinear problems 250–252 11, 12, 15, 165
solver 227–229, 230–232 Min-max scaling 40, 59–61
with two variables using graphs Mixed-integer optimization models
220–227 284–285
Linear regression 23, 25, 55–59 Monotonic constraint
and least squares 51–52 integration
multiple linear regression 53–55 into artifical neural network
logistic regression 56–59 (ANN) 186–190
polynomial regression 55–56 into XGBoost 184–186
shrinkage 59 MSE. See Mean-squared error (MSE)
LASSO regression 61–62 Multidimensional scaling (MDS) 35
ridge regression 60–61 Multiple linear regression 53–55
logistic regression 56–59
polynomial regression 55–56
Machine learning (ML) modeling 23, Mutual information (MI) 42
26–28
to address maritime transport
problems 29, 31, 32 Naive Bayes classifier 63–68
assessment 50 Naïve-Bayes Decision Tree (NBTree) 69
304  Machine learning and data analytics for maritime studies

Natural integer decision variables Rand index (RI) 150–151

277–280 Random forest (RF) 105, 118–119, 186,
Network flow optimization 233–247 203, 207
Neuron 93–94 Receiver operating characteristic (ROC)
New inspection regime (NIR) 12, curve 45–46
13, 182 Rectified linear units (ReLU) activation
function 97
Recursive feature elimination (RFE) 42
Objective function 211, 212, 215 Remote follow-up inspection 12
Ocean freight market condition Ridge regression 23, 25, 60–61
prediction 168–169 Root-mean-squared error (RMSE) 43
Oil tanker 2
One-hot encoding 40, 54, 85–86, 101
Orient Overseas Container Line Safety management 9–10
(OOCL) 5–6 Scikit-learn API 52, 54, 55, 59–61, 85,
Origin-destination (OD) pair 7 88, 91, 101, 112, 114, 119, 126,
Out-of-bag error 48 158, 200
Semi-naive Bayes classifiers 68–73
Shapley additive explanations (SHAP)
Paris MoU 11, 12, 14, 15 200, 202–208, 300
Pearson’s correlation coefficient Ship accident data 34
(Pearson’s r ) 41 Ship broker 5
Polynomial kernel 88, 89, 91 Ship destination port and arrival time at
Polynomial regression 23, 25, 55–56 port prediction 169
Port condition prediction 169 Ship energy efficiency prediction
Port state control (PSC) 168
memorandum of understanding Ship management company 3
(MoU) on 11 Ship operator 3
ship inspection 10 Ship owner 3
background 11 Shipping industry
data set 15–18 key roles
development 11 charterer 4
onboard procedure 12 classification society 4
results 14–15 flag state 3–4
ship selection 11–14 freight forwarder 4
Port state control officer (PSCO) 11 ship broker 5
Port statistics 34 ship management company 3
Posthoc explanation 199 ship operator 3
Precision 44–45 ship owner 3
Predictive accuracy (PDR) 196 Ship risk prediction
Predictive problem 21–23 feature monotonicity 181
Principal components analysis (PCA) oveview 182–184
35, 40 safety management 167–168
PyDotPlus API 200 Ship risk profile (SRP) 12, 165
 Index  305

Ship sailing records 33–34 Tanh activation function 96

Ship trajectory prediction 167 Target encoding 40
Shrinkage linear regression models Tokyo MoU 11–15, 18, 101, 139, 140,
59 171, 172, 174, 182–184, 204–206
LASSO regression 61–62 Tree augment naive Bayes (TAN)
ridge regression 60–61 classifier 68, 70, 71, 73
Sigmoid activation function 95 Tree-based models
Sigmoid kernel 89 decision tree 105–106
Simple linear regression 51–52 ensemble learning 116
Single-linkage clustering 148–149 Adaboost 122–125
Soft margin support vector machine bagging 116–221
83–86 boosting 121–130
Spearman’s rank coefficient node splitting
(Spearman’s r) 41 in classification decision tree
Standardization 40 (see Classification decision tree)
Statistical modeling 23 in regression trees 113–117
Stepwise regression 23, 25
Support vector machine (SVM) 79
hard margin 79–83 Unweighted average-linkage clustering
Kernel trick 86–89 149
soft margin 83–86
support vector regression 90–92
Support vector regression (SVR) 79, World fleet structure
90–92 bulk carrier 1
Synthetic minority oversampling container ship 2–3
technique (SMOTE) 59 oil tanker 2
Applications of Machine Learning and Data Analytics
Models in Maritime Transportation

Applications of Machine Learning

Machine learning and data analytics can be used to inform technical, commercial and financial

Transportation
Analytics Models in Maritime
Applications of Machine Learning and Data
decisions in the maritime industry. Applications of Machine Learning and Data Analytics
Models in Maritime Transportation explores the fundamental principles of analysing maritime-
transportation related practical problems using data-driven models, with a particular focus on
machine learning and operations research models.
and Data Analytics Models in
Data-enabled methodologies, technologies, and applications in maritime transportation are
clearly and concisely explained, and case studies of typical maritime challenges and solutions
Maritime Transportation
are also included. The authors begin with an introduction to maritime transportation, followed
by chapters providing an overview of ship inspection by port state control, and the principles
of data driven models. Further chapters cover linear regression models, Bayesian networks,
support vector machines, artificial neural networks, tree-based models, association rule
learning, cluster analysis, classic and emerging approaches to solving practical problems in
maritime transport, incorporating shipping domain knowledge into data-driven models,
explanation of black-box machine learning models in maritime transport, linear optimization,
advanced linear optimization, and integer optimization. A concluding chapter provides an
overview of coverage and explores future possibilities in the field.
Ran Yan and Shuaian Wang
The book will be especially useful to researchers and professionals with expertise in maritime
research who wish to learn how to apply data analytics and machine learning to their fields.

About the Authors

Ran Yan is a research assistant professor in the Department of Logistics and Maritime Studies
at The Hong Kong Polytechnic University, China.

Shuaian Wang is a professor in the Department of Logistics and Maritime Studies at The
Hong Kong Polytechnic University, China.

Yan and Wang

The Institution of Engineering and Technology

theiet.org
978-1-83953-559-8