Accepted Manuscript

Three-way decision and granular computing

Yiyu Yao

PII: S0888-613X(18)30280-9
DOI: https://doi.org/10.1016/j.ijar.2018.09.005
Reference: IJA 8255

To appear in: International Journal of Approximate Reasoning

Received date: 15 April 2018

Revised date: 18 July 2018
Accepted date: 13 September 2018

Please cite this article in press as: Y. Yao, Three-way decision and granular computing, Int. J. Approx. Reason. (2018),
https://doi.org/10.1016/j.ijar.2018.09.005

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing
this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is
published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
Highlights
• Discuss a wide sense of three-way decision as thinking in threes.
• Propose a TAO (trisecting-acting-outcome) model of three-way decision.
• Explore the interplay of three-way decision and granular computing in the context of cognitive science.
• Explain granular computing according to a philosophy of thinking in threes.
• Illustrate the value of three-way decision by many examples.
 Three-way decision and granular computing
Yiyu Yao
Department of Computer Science, University of Regina,
Regina, Saskatchewan, Canada S4S 0A2
email: [email protected]

Abstract
Based on results from cognitive science, this paper examines the two fields
of three-way decision and granular computing, as well as their interplay.
The ideas from one field shed new light on the other field. The integration
of the two gives rise to three-way granular computing, that is, thinking,
problem solving, and information processing in threes. We discuss a wide
sense of three-way decision and propose a trisecting-acting-outcome (TAO)
model. We explain fundamental notions of granular computing based on
the philosophy of three-way decision as thinking in threes. We discuss a
model of three-way granular computing by making use of two particular
types of granular structures represented, respectively, by three granules and
three levels. We use examples across different disciplines to demonstrate the
values of the two types. Our investigation suggests that, in many situations,
the power of granular computing is indeed the power of three-way decision,
i.e., thinking in threes.
Keywords: Three-way decision; three-way computing; granular computing
in threes; thinking in threes; magical number three

1. Introduction
A recently proposed theory of three-way decision (3WD) [74, 76, 79] pro-
motes thinking, problem solving, and information processing in threes, that
is, using three parts, three elements, three components, three perspectives,
three views, three levels, three generations, three periods, three stages, three
steps, triangles, triads, triplets, and many others. A narrow sense of three-
way decision [43, 79] was introduced for interpreting the three types of clas-
sification rules in rough set theory [49]. In order to apply three-way decision

Preprint submitted to Elsevier September 17, 2018

to a wide spectrum of domains, we need to consider a wide sense [74], in
which “decision” may be viewed as computing, processing, analysis, and so
on. We can take either a literal sense of “three,” as being taken in the present
study, or a figurative sense of “three” as “a few” or “about three.” It is this
wide sense that increases the power of three-way decision.
In explaining a philosophy of thinking in threes, Minsky says1 , “I find
most people say, well it’s either this or that and I’m always inclined to look
for a third thing. ... if somebody says, is it left or right, I’m always looking
for a third way. ... of course, most of the time you can’t find one, but every
now and then I get a new theory ... .” A useful lesson from Minsky is trying
to think in three whenever you have two, although it may fail sometimes.
This is just one of many examples to be examined to demonstrate the needs
for and the benefits of three-way decision as thinking in threes.
In the context of cognitive science, in this paper we study the interplay of
three-way decision and granular computing as two specific human-inspired
paradigms of cognitive computing. We discuss a wide sense of three-way
decision and propose a trisecting-acting-outcome (TAO) model. We explain,
in the light of three-way decision, the basic questions, notions and concepts of
granular computing, leading to a conception of granular computing in threes
or three-way granular computing.

1.1. Three-way decision and granular computing in the cognitive era

Categorization and abstraction are basic cognitive tools for dealing with
the complexity of the real world. We group ideas, objects, and events into
categories, usually for a specific purpose, by recognizing their similarity, dis-
similarity, or functionality. The ideas of categorization and abstraction serve
as a cognitive basis for both three-way decision and granular computing.
Three-way decision is about thinking, processing, and computing in threes.
It is motivated by an observation that using three parts, components, and
levels to understand a whole appears to be practiced universally in human
problem solving. We can relate three parts of three-way decision to the ideas
of categorization and abstraction, namely, three-way decision uses three cat-
egories of things. Granular computing concerns theory and practice of struc-
tured multilevel approaches, in which each level of granularity or abstraction
is represented and interpreted through a family of granules. We can draw

1
https://www.youtube.com/watch?v=_POW-geD5no

2
some correspondence between granules used in granular computing and cat-
egories used in cognition and human understanding. If we further draw a
correspondence between three parts used in three-way decision and granules
used in granular computing, we can unify the two fields of studies as special
types of human-inspired paradigms of computing.
Kelly [34] puts forward a three-era evolution model of computing, consist-
ing of the tabulating era (1900s-1940s), the programming era (1950s-onward),
and the cognitive era (2011- ). The cognitive era is marked by a shift of at-
tention towards human-machine cooperation, integration, and inter-creation.
Studying human ways to problem solving, human-inspired algorithms and
systems, and human understandable systems is of significant importance in
the cognitive era. As exemplified by Kelly’s three-era classification, granular
thinking in general and three-way thinking in specific are everyday common
human practice in complex problem solving. The view of three-way decision
and granular computing as two special models of cognitive computing is the
basis of the investigation of the present paper.

1.2. Three basic questions of granular structures

Granular computing explores effective use of granules and granular struc-
tures constructed from granules. Granules, having a similar interpretation as
categories, are clusters, groups, and sets of ideas and objects. Although basic
ideas and principles of granular computing have appeared frequently across
many disciplines, a study of granular computing as a new scientific field of
investigation [7, 20, 50, 52, 53, 64, 71, 78] is much influenced by artificial
intelligence [26, 77], fuzzy sets [84], and rough sets [49, 81].
The triarchic theory of granular computing consists of three main compo-
nents: philosophy of structured thinking, methodology of structured problem
solving, and mechanism of structured information processing [73]. A key un-
derlying notion is a granular structure characterized by granules, levels, and
a hierarchy [80]. Granules are units, focal points, and vocabularies for repre-
senting and interpreting a problem. An essential feature of granules is their
granularity or degree of abstraction, which may be intuitively interpreted in
terms of the size of granules. A level of granularity or abstraction is formed
by granules of similar nature or granularity. Levels are (partially) ordered
according to their granularity, resulting in a hierarchy of multiple levels.
When constructing and using granular structures, one may ask at least
the following three fundamental questions:

3
(1) Are the granules, levels, and hierarchy meaningful for a particular prob-
lem?

(2) What are the most appropriate number of levels and the number of
granules?

(3) What is the most effective level of granularity?

Some attempts have been made to address these questions. The principle
of justifiable granularity, proposed by Pedrycz [51], represents an important
initiative in answering question (1). Several authors discuss question (2)
by drawing attention to Miller’s [47] influential work on limited information
processing capacity of humans, which is about 7 ± 2 units of information.
Yao [80] considers its implications to constructing granular structures. Chen
and Du [13] discuss its uses in economic decision making in an interdisci-
plinary context. One of the conclusions is that it is effective and sufficient to
use only a few granules or a few levels. In many data analysis models, differ-
ent subsets of attributes or features induce possibly distinct granular views of
the same data set. Regarding question (3), attribute reduction in rough set
theory [49] and feature selection in pattern recognition and machine learning
may be viewed as a search for an appropriate level of granularity.

1.3. Main results and the organization of this paper

A study of three-way decision and granular computing, in the context of
results from cognitive science, is by itself an example of thinking in threes.
Figure 1 is a conceptual map of the investigation of this paper. We discuss
pertinent ideas, notions and concepts for each of the three fields, as indicated
by the rectangular boxes at the three corners of the triangle. The numbers
inside the brackets give the sections or subsections in which materials about
the respective field are covered. For examples, Sections 2, 3 and 5 contain
materials about three-way decisions. We study pair-wise relationships of
the three fields. We examine, in particular, three types of relations: a) one
field provides a basis of another field as indicated by solid lines labeled by
“support,” b) one field provides a special model for another field as indicated
by dotted lines labeled by “special model,” and c) methods and tools in one
field can be used to explain ideas of another field as indicated by dashed lines
labeled by “explanation.” The numbers inside the brackets following a label
are the sections or subsections in which the particular relation is discussed.
For example, cognitive science provides a basis for three-way decision, which

4
 Three-way Decision
(2; 3; 5)

)
.3 su

.2

sp
)
.2

ex
;2

(3
pp

ec
(3

pl
.2

ia
el
or

an
n
(2

lm
od
t

io

at
t (5

at
lm
or

od
io
.1

an

n
p

el
ia
)
up

pl

(4
ec

(5
ex
s

)
sp

)
explanation (3.2)

Cognitive Science special model (3.2) Granular Computing

(2.2; 2.3; 3.2) support (2.2) (4; 5)

Figure 1: Relations among cognitive science, three-way decision, and granular computing

is discussed in subsections 2.2 and 2.3. The philosophy and methodology

of three-way decision, as thinking in threes, can be used to study granular
computing, which is discussed in Section 4.
By using the structure of Figure 1 as a blueprint, we demonstrate that
answers, or at least partial answers, to the three basic questions of granular
structures may be found in a theory of three-way decision [74]. To achieve
our goal, we organize the rest of the paper into three parts as follows.
Part I consists of Sections 2 and 3. In Section 2, we present a wide
sense of three-way decision. The magical number three underlies three-way
decision and may hold the key to realize the power of granular computing. We
examine three narrow senses of three-way decision. We look at the cognitive
basis of thinking in threes, as well as its optimality. In Section 3, we present a
trisecting-acting-outcome (TAO) model of three-way decision. We use Parr’s
captivology [48] as an example to illustrate the working of the TAO model.
Part II consists of Section 4. We explain many notions of granular com-
puting under the light of three-way decision. Three-way decision corresponds
to granular computing in threes. We explain granular computing from the
three perspectives of philosophy, methodology, and mechanism. We explain
granular structures in terms of the three components of granules, levels,
and hierarchies. We consider the three modes of top-down, bottom-up, and
middle-out computing.
Part III consists of Section 5. We use “granular computing in threes” or
“three-way granular thinking” to label a special class of models resulted from

5
an integration of three-way decision and granular computing. It is a model
of thinking in threes, problem solving in threes, and computing in threes.
In many situations, using three parts, three granules, or three levels plays a
crucial role in a problem-solving process. We examine two types of granular
structures that make use of threes. One type consists of structures with
three levels and multiple granules at each level, which is called the 3LmG
type. The other type consists of structures with multiple levels and three
granules at each level, which is called the mL3G type. We use examples across
many disciplines to demonstrate the value of the two types of structures for
thinking, problem solving, and information processing in threes.

2. A wide sense of three-way decision

Three-way decision is thinking in threes and computing in threes, where
three is the number of parts, elements, and components under consideration.
We examine both a narrow sense and a wide sense of three-way decision,
present three specific senses of three-way decision as thinking in threes, and
discuss the cognitive basis of thinking in threes.

2.1. Three-way decision as thinking in threes

In rough set theory [49], a set is approximated by three pair-wise disjoint
sets called the positive, boundary, and negative regions. The notion of three-
way decision (3WD) is introduced to interpret the three types of decision
rules obtained from the three regions in a probabilistic rough set model [79],
namely, acceptance, non-commitment, and rejection rules. The earlier re-
search mainly focused on this narrow sense of three-way decision [43]. Most
recent studies have led to a wide sense of three-way decision [74]. The wide
sense of three-way decision is built on a philosophy of thinking in threes. To
have a full understanding of the wide sense, we have at least the following
three related but distinct specific senses.
A third. We interpret three-way decision as trichotomous thinking,
which introduces a third option to dichotomous thinking with two options.
With three-way decision, we move from true/false, black/white, and yes/no
to true/unsure/false, black/grey/white, and yes/maybe/no. The third op-
tion provides the necessary flexibility and universality of thinking in threes.
Trisection. We interpret three-way decision in terms of the notions of
parts and whole, as given in a trisecting-and-acting (T&A) framework of
three-way decision [74]. The basic ideas are (a) to divide a whole into three

6
parts, and (b) to devise the most effective strategies to act upon the three
parts. Dividing a whole into three parts reduces the complexity of the whole
and leads to simplicity.
Triad. We interpret three-way decision with reference to a common prac-
tice of using a triad or triplet consisting of three elements or three compo-
nents. A triad offers sufficient complexity in terms of individuals and relation-
ships between them and, at the same time, is simple enough for memorization
and processing.
All three specific senses use triplets of three options, parts, or elements.
They differ with respect to the semantics of integrative triplets. An example
of the first specific sense is three-valued logic in which a third value is used
to denote a value different from both true and false. An example of the
second specific sense is to divide tax payers into low, middle, and high income
classes and devise corresponding strategies for each class. An example of
the third specific sense is the use of three words, three phrases, and three
sentences in writing and speech. In many cases, the three specific senses
may be independent, may introduce each other, and may appear together.
For example, a trisection may introduce a third option and a triad of three
parts. A triad may consider three options without a clear reference to the
process from which the triad is derived. The possible combinations of the
three specific senses offer a desirable flexibility of three-way decision. Our
interpretations of three-way decision, in terms of three specific senses of “a
third,” “trisection,” and “triad,” are indeed an example of thinking in threes.
In other words, we divide the whole of three-way decision into three parts
represented by the three specific senses.
In a set-theoretic setting, a trisection of a universal set consists of three
pair-wise disjoint subsets such that their union is the universal set. There are
two other equivalent representations of a trisection. One representation uses
a pair of nested sets, such as the pair of lower and upper approximations in
rough sets [49] and the pair of lower and upper bounds in interval sets [72].
The lower approximation/bound forms one set, the difference between the
upper approximation/bound and the lower approximation/bound forms an-
other set, and the complement of the upper approximation/bound forms the
third set. The other representation uses a pair of disjoint subsets called an
orthopair, which was proposed and systematically studied by Ciucci [15, 16].
The third subset in the trisection can be obtained by the set complement of
the union of the two subsets in the orthopair.
The wide sense of three-way decision offers a new understanding. By fo-

7
cusing on “three-way” as the use of threes, we may replace “decision” by other
words to introduce new types of three-way approaches, such as three-way
thinking, three-way computing, three-way processing, three-way classifica-
tion [10, 37, 83, 87, 88], three-way analysis, three-way clustering [1, 82], three-
way recommendation [5, 85], three-way decision support [67, 70], three-way
concept analysis [57, 59, 63, 72], three-way concept learning [28, 38], tempo-
ral and spacial three-way decision [42], three-way attribute reduction [45, 86],
and many more [27, 39, 40, 44]. Studies up to date indicate that a theory of
three-way decision in the wide sense is urgently needed, comes at the right
time, and will rise in the future.

2.2. Cognitive basis of thinking in threes

One plausible justification for using threes in three-way decision is our
limited capacity of short-term working memory, a fundamental concept in
cognitive science and cognitive psychology. We can only process a small
number of units of information in short-term memory. The use of the magi-
cal number three in three-way decision is consistent with and supported by
results from cognitive science.
In 1956, Miller [47] reported an important finding on the capacity of
human information processing in short-term memory. We can normally hold
7 ± 2 units of information in short-term memory for immediate recall and
processing. A clever solution to this limitation is a technique known as
chunking. Through encoding, a large amount of information is chunked into
a manageable number of units or chunks. Simon [62] suggested the magical
number five as the number of chunks, instead of Miller’s magical number
seven. Later studies, for example, Cowan [18] and Gobet and Clarkson [24],
indicated that the actual number of chunks is smaller than seven, is about
or below four, and is three or two. Existing studies show that there does not
exist a general agreement on the exact number of chunks. The disagreements
may be explained by the fact that the actual magical number, if exists at all,
depends largely on the interpretation of chunks and the size of chunks. But
we do not have a common agreed understanding of the latter. Nevertheless,
there is a general agreement: we can only manage a small number of chunks
and this number can be qualitatively described as “a few,” which is usually
less than seven.
Warfield [69] presented a very interesting argument that relates the num-
ber three and the number seven. The argument considers the interaction
between chunks. Let A, B, and C denote three chunks. If we consider them

8
as individuals, we have three chunks A, B, and C. If we consider the pair-
wise relationships, we have three pairs (A, B), (A, C), and (B, C). Finally,
if we consider all three chunks together, we have one triplet (A, B, C). The
total number is 3 + 3 + 1 = 7, which is exactly magical number seven. By
following the same argument, if we consider two chunks A and B, we have
two individuals and one pair (A, B). It gives rise to a total of 3. That is, we
obtain the magical number three from two chunks. If we use four chunks,
we would have four individuals, six pairs, four triplets, and one quadruple.
The total becomes 4 + 6 + 4 + 1 = 15, which is far more beyond the ca-
pacity as suggested by the magical number seven. An important implication
of Warfield’s argument is that when relationships between chunks are also
considered, the magical number is three, which is the one used in three-way
decision and three-way computing.

2.3. Economy of the base three system

An additional piece of evidence in support of thinking and processing in
threes is the economy of the base three system and ternary numbers [21, 25,
55].
Given a number (an−1 . . . a0 )b of base b in a positional notation, its decimal
value is given by:
an−1 bn−1 + · · · + a0 b0 , (1)
where 0 ≤ ai ≤ b − 1, i = 0, 1, . . . , n − 1. The cost of a system of a particular
base is determined by two factors. The base b is the depth or radix of a
number, which is the number of different symbols in each digit position. The
number of digits in a number is the width of the number. As suggested
in [21], bn represents “a fair estimate of the number of tubes required in the
system.” If n digits are used, the number of numbers expressible is bn , that
is, 0, 1, . . ., bn − 1. By holding bn as a constant, that is, the number of
numbers to be expressed is a constant, one can minimize bn to find the most
economical base b. If one treats both b and n as continuous variables, the
problem can be easily solved. The optimal value is e, the base of the natural
logarithm [21, 25], that is, b = e = 2.71828 . . .. Hayes [25] shows that for
integer values of b and n, 3, the integer closest to e, is “almost always the
most economical integer radix.”
As an interesting application of ternary numbers, Hayes suggests the use
of third-cut file folders, i.e., folders in which “the tabs appear in three po-
sitions, left, middle, and right,” as an alternative to half-cut folders. With

9
third-cut folders, you can easily insert a new folder that differs from its two
neighbors. Furthermore, base three is the smallest base having this prop-
erty. The third-cut folders are a simple example to demonstrate the power
of thinking in threes.
Phythian [55] presents another view on the economy of number bases for
the purpose of reading and understanding numbers, instead of representing
and implementing numbers by physical devices. With respect to the depth b,
we need b symbols, which has the same interpretation as before. With respect
to the width n, we need n symbols for b0 , b1 , . . . , bn−1 , which is different from
the earlier interpretation. For example, in the decimal system, we have one
for 100 , ten for 101 , and hundred for 102 . It follows that for a number of n
digits, we need a total of b + n symbols, assuming that the different symbols
are used for b0 , b1 , . . . , bn−1 (as we will see later, this may not be true in
general). By holding bn as a constant, we can minimize b + n to find the
most economical base. By treating b and n as continuous variables, one can
find the most economical base by solving the equation b(loge b)2 = loge N ,
where N is the number of numbers to be expressed. In this case, the optimal
base depends on the value of N .
According to Phythian’s argument, with integer bases and integer num-
bers, base ten is the most economical base in the sense that it only requires 34
symbols for numbers from 0 to 1024 . For numbers from 0 to 1010 , base seven
is the most economical base that requires 18 symbols, while base 10 requires
20 symbols. The results seem to suggest that if we only use numbers smaller
than 1010 , which is actually the case in reality, we perhaps should reduce our
base. However, the actual practice is very different: we use decimal system
rather than base seven system. This seemingly inconsistency needs a closer
examination.
In defining the economical index r + n of base b system with n dig-
its, Phythian assumes that distinct symbols are required for all n positions,
leading to a conclusion of base reduction. As an alternative solution, we
may choose to name some positions rather than all positions. This is, in
fact, the solution offered by a clever way of counting by threes, in which the
magical number three plays an important role. We name positions at two
levels. At a lower level, we have one for 100 , ten for 101 , hundred for 102
as the three basic names. At a higher level, we have one for 100 , thousand
for 103 , million for 106 , and billion for 109 , trillion for 1012 , and so on with
an increment of three in the power of ten. To explicitly express counting by
threes, we insert a comma to group digits in a number into groups of threes.

10
For example, 1200000 is written as 1,200,000. This immediately provides a
two-level understanding of the magnitude of numbers. The high level, indi-
cated by commas, is given by ones, thousands, millions, and so on. For a
specific higher level magnitude, we can refine it into three refined magnitude
by using ones, tens, and hundreds to read the three numbers for that mag-
nitude. For example, at the magnitude of millions, we have millions, tens of
millions, and hundreds of millions. In this way, 1,200,000 is one million and
two hundred thousands. Through counting by threes, for numbers from 0
to 1012 , we actually use 17 symbols, which is less than 18 symbols required
by the base seven system. That is, the decimal system is actually the most
economical system with the technique of counting by threes. This is another
example to illustrate the power of thinking in threes.
As an interesting application of Phythian’s model, we can justify the use
of base three system when a small number of numbers is under consideration.
An ancient use of base three system was introduced by Chinese scholar Yang
Xiong (53 BCE-18 CE) for coding 81 numbers in Tai Xuan Jing (The Canon
of Supreme Mystery). Based on the formulation given by Phythian, for
81 numbers, base three, base four, and base five systems are all the most
economical systems that require seven names. Similar to the earlier example
of third-cut folders given by Hayes [25], base three is the smallest base for
coding 81 numbers most economically from an understanding point of view.
This example demonstrates again the power of thinking in threes.

2.4. Further discussions

If we accept the hypothesis that the capacity of short-term memory, mea-
sured in chunks independent of the materials for forming a chunk, is three, we
should find many examples of thinking in threes in real life. This is indeed the
case. Thinking in threes is a frequently occurring theme in human knowing,
understanding and problem solving [9, 36, 74], for example, three parts of a
whole, three components of a system or a theory, three perspectives/angles of
an issue, three levels of understanding, etc. Languages, as a basic cognitive
tool for thinking, understanding, and communication, provide an excellent
example to illustrate that three is a magical number in thinking. Take En-
glish as an example. We have three degrees for adjectives and adverbs, that
is, positive, comparative, and superlative. We use three basic tenses of verbs,
that is, past, present, and future tenses. For representing points of view, we
may use any one of three choices, namely, the first person, the second person,
and the third person. Furthermore, we frequently use triplets, that is, three

11
words, three phrases, and three sentences, in writing and speech. In gen-
eral, approaches that use threes appear across many fields and disciplines,
including computer science, information science, statistical science, manage-
ment science, engineering, social science, medical decision-making, and many
more. Our limited information processing capacity dictates us to think in
threes. Thinking and processing in threes are a kind of our second nature
and have been widely practiced in everyday life.
The two interpretations of the economy of numerical bases shed light
on the use of the magical number three. The product of the depth and
width, bn, is the cost of representing numbers with physical devices. The
sum of the depth and width, b + n, is the cognitive cost in understanding
and reading numbers. Although the magical number three appears in both
interpretations, it might not be the case in general. In developing a theory
of three-way decision, we focus more on the cognitive perspectives. In other
words, we explore the economy of three-way decision to reduce cognitive
overload. The method of counting by three in reading decimal numbers
is an excellent example to illustrate our position. By grouping digits in a
number into groups of threes, it is much easier for us to read, interpret, and
use decimal numbers. The method does not reduce the number of physical
devices in representing a number, which requires the use of base three. On
the other hand, using base three system will lead to a cognitive difficulty,
as we do not have names for 30 , 31 , . . .. Therefore, thinking in threes with
decimal numbers is a much preferred choice.
The wide sense of three-way decision as thinking in threes has a solid cog-
nitive basis and is supported by a vast amount of evidence. We may formulate
models of three-way decision based on everyday uses of three parts, three ele-
ments, three components, three perspectives, three views, three levels, three
generations, three periods, three stages, three steps, triangles, triads, triplets
and many others. Thinking in threes avoids the oversimplification of thinking
in ones and twos and, at the same time, does not suffer from the complexity
of thinking in fours or more. In developing a theory of three-way decision,
the magical number is three [74].

3. A TAO model of three-way decision

According to the principle of thinking in threes, we propose a trisecting-
acting-outcome (TAO) model by adding a third component to the original

12
trisecting-and-acting model [74]. We use Parr’s captivology [48] to explain
the working of the TAO model.

3.1. The model

A central idea of three-way decision is thinking in threes. To model the
specific “trisection” sense of three-way decision, we consider three related
questions:

(1) How to construct a trisection from a whole?

(2) How to devise strategies to process the three parts of the trisection?

(3) How to evaluate the effectiveness of the trisection and strategies?

An earlier trisecting-and-acting model [74] only covers the first two ques-
tions. To give a more complete picture, we add a third component about the
evaluation. This leads to a trisecting-acting-outcome (TAO) model of three-
way decision with three components. Existing studies on three-way decision
concentrate on trisecting and acting. Only a few papers touch upon the issue
of outcome evaluation [23, 31].
Figure 2 presents a high-level description of the TAO model with three
components of trisecting, acting, and outcome evaluation. The function of
trisecting, denoted by solid lines, is to divide a whole into three related and
relatively independent parts. The resulting three parts are called a trisection
of the whole. The function of acting, denoted by dashed lines, is to apply a
set of strategies to process the three parts. By trisecting a whole and acting
on the resulting trisection, it would produce an expected outcome. The func-
tion of outcome evaluation is to measure the effectiveness of the results from
a combined effort of trisecting and acting as enclosed by the large dashed
rectangular box in the figure. The TAO model provides an architectural
framework of three-way decision. When applying the model to a particular
application, one needs to use semantically and physically meaningful trisec-
tions, profitable actions, and informative measures of effectiveness.
There are many ways to divide a whole into three parts and many strate-
gies to act upon the three parts. We may have strategies for a particular part,
strategies for two parts, and strategies for all three parts, respectively. The
effectiveness of three-way decision depends on a proper match of trisecting
methods and strategies for actions. Several modes of combination are possi-
ble. The simplest mode assumes the independence of trisections and actions,

13
 Whole

Trisecting

Trisection-driven 3WD
Action-driven 3WD

Part I Part II Part III

Acting

Strategies

Outcome evaluation

Figure 2: TAO model of three-way decision

as was done in earlier studies of three-way decision. This is usually done

by considering the trisecting methods and strategies separately. In Figure 2,
we give two trisection and action interdependent modes. A trisection-driven
mode requires that actions and strategies are built based on results of tri-
sections. In other words, a trisecting method guides the search for effective
strategies and their applications. Conversely, an action-driven mode starts
with a set of strategies and a suitable trisection is sought so that these strate-
gies can be effectively applied. In other words, strategies determine how to
trisect a whole. In many situations, a combined mode of trisection-driven
and action-driven may be preferred. It can be a co-supportive iterative pro-
cess in the sense that a trisection provides hint for devising better strategies
and a set of strategies may further improve trisecting. The iterative circle
continues until we arrive at a right combination of a trisection and a set of
strategies.

3.2. An illustration of the model

The high-level description of the TAO model of three-way decision enables
us to explain many commonly used three-way approaches in human problem
solving. We use a simple practical example to explain the working of the
model.
In studying human memory, a three-way classification scheme is typically
used, that is, we have three types of memory known as sensory memory,
short-term memory, and long-term memory [4, 56]. According to Plotnik

14
and Kouyoumdjian [56], sensory memory is a process of receiving and hold-
ing environmental raw information for only a brief period of time, short-
term memory is a process of holding a limited amount of information for
only a short period of time, and long-term memory is the process of holding
large amounts of information over long periods of time. Each type of mem-
ory serves for specific functions, sensory memory for recording, short-term
memory for working, and long-term memory for storing. Furthermore, selec-
tive attention transfers sensory memory to short-term memory and encoding
transfers short-term memory to long-term memory. To a large extent, human
problem solving is constrained by the three types of memory. Correspond-
ing to the three types of memory, there are different cognitive strategies and
tools.
Studying human memory in three types is an example of thinking in
threes. The three-way classification is meaningful and is supported by ample
empirical evidence. The division of the three types gives a clear and simple
understanding of memory in terms of both time periods and the associated
functions. It might be possible to build another model that has more than
three types. The model would be unnecessarily complicated in the light of
the hypothesis of a limited capacity of short-term memory. Given a trisection
of memory, an important task is to examine strategies that make effective
use of the three types of memory.
In developing a theory and practice of captivology, Parr [48] considers
three types of attention, namely, immediate attention, short attention, and
long attention. We can easily establish a correspondence between the three
types of memory and three types of attention. Two essential tasks of captivol-
ogy are capturing the three types of attention and facilitating the transfer of
immediate attention to short attention and short attention to long attention.
In the light of the TAO model of three-way decision, captivology focuses on
action and outcome according to the trisection of attention.
Based on features and functions of the three types of attention, Parr
examines a set of strategies and gives real world examples to demonstrate
their effectiveness. As examples for illustration, we briefly mention some of
these strategies. Immediate attention is an automatic response, working in
a similar way as sensory memory. To capture immediate attention, one may
explore strategies of creating contrast and association. In other words, one’s
immediate attention will be drawn by something that is very different from its
surroundings. Building associations is a basic subconscious cognitive activity
and associations lead to attention. To transfer immediate attention to short

15
and long attention and to retain long attention, many different strategies
may be used. For example, people pay attention to things that fit their
frame. It is effective to adapt to the audiences’ frame of reference to capture
their attention. It is also possible to explore the scarcity effects, bizarreness
effects, and mystery effects, as all of them are closely related to gaining
attention. Using rewards, including, extrinsic and surprising rewards, as well
as acknowledgements, proves to be effective attention grasping strategies.
Finally, in order to have long attention, it is crucial to create values and to
build a reputation. People will remember things or people by their values
and reputation.
Many of the strategies given by Parr [48] are applicable to two or three
types of attention, rather than one particular type. Some strategies facilitate
moves from one type to another type. This is related to a specific model of
three-way decision known as actionable three-way decision [23].

4. Granular computing explained in threes

By applying the principles of thinking in threes, we explain the basic
concepts and notions in granular computing in threes, namely, the three
elements of granular structures, the triarchic theory of granular computing,
and the three modes of granular information processing.

4.1. Three elements of granular structures: granules, levels, and hierarchies

Granular computing concerns structured approaches to thinking, problem
solving, and information processing [7, 8, 20, 22, 26, 52, 53, 64, 68, 71, 73,
77, 78]. The power of granular computing is the power of useful structures.
Granular structures should reflect three important aspects of structuredness.
They are called the three M’s of granular structures, namely, multilevel,
multiview, and multipurpose. Granular structures provide a multilevel un-
derstanding from a particular view, offer a multiview vision covering various
angles, and have multiple functions to serve different purposes. While the
first two M’s have been extensively discussed [78], the third M is introduced
in this paper to emphasize the role and effectiveness of granular structures.
We explain granular structures based on three main ingredients: granules,
levels, and hierarchies [73, 77, 80, 81].
Granules. In granular computing, granules may be interpreted as the
units, elements, concepts, and notions that we use for representing and de-
scribing a complex system or a complex problem as a whole. For example, in

16
systems theory we may interpret granules as the parts of a whole [11]. In hu-
man information processing, information granules correspond to the notion
of information chunks. That is, a chunk is a granule representing a mean-
ingful whole that binds together individual pieces of information. In text
processing, granules at various levels are words, sentences, paragraphs, and
articles, or alternatively, titles, section headings, and subsection headings.
In different contexts, we use different languages and vocabularies to define
and explain granules. There are two important characteristics of granules.
One is the granularity of granules, which may be considered as the size of
granules. The other is the dual roles played by granules. A granule may be a
whole consisting of many granules as its parts. At the same time, the same
granule may serve as one part of another granule. According to these two
features, we can order granules according to their granularity and we can
also connect granules to form a web of granules. This enables us to construct
very general granular structures in a graph-theoretic setting. In the web of
granules, regions of the web give rise to another type of granules.
Levels. Although a web of granules has the desired generality and ex-
pressive power, in many practical situations we only use relatively simple
granular structures. A hierarchical granular structure consists of multiple
levels. Each level is represented and characterized by a family of granules of
similar nature and similar granularity. The levels are ordered according to
the granularity of granules involved. A higher level controls and determines
its next lower level. We add more detailed information and more concrete
ideas as we move top-down. A lower level supports its next upper level. We
remove particular information and abstract the essential ideas as we move
bottom-up. With a hierarchy, we have a multilevel understanding, which
also leads naturally to multilevel processing.
Hierarchies. A hierarchy is a representation of reality from a particular
angle, rather than the reality itself. An advantage, and at the same time a
disadvantage, of a representation is that it makes certain aspects explicit and
clear at the expense of pushing some other aspects to the background [46].
Furthermore, one representation may only serve the purpose of its intended
applications and will not serve the purpose of other applications. To com-
pensate for the shortcomings of a particular hierarchy and to serve multiple
purposes, we use many hierarchies for a multiview vision.
A granule at each level provides a partial, fragmentary understanding. A
family of granules forms a level and provides a full, complete understand-
ing at a particular level of granularity. A family of levels forms a hierarchy

17
and provides a multilevel understanding from a specific viewpoint. Finally,
a family of hierarchies provides a multiview understanding from many view-
points. That is, granular computing is both a multilevel and a multiview
approach. While multilevel accounts for different levels of abstraction and
details, multiview avoids the potential pitfalls caused by bias of a single view.
The first two M’s of granular structures, i.e., multilevel and multivew,
state the desired structural properties, as captured by the notions of granules,
levels, and hierarchies. The third M of granular structures, i.e., multipurpose,
states the desired functional property. As a matter of fact, the first two M’s
naturally offer the third M, that is, the multiplicity of levels and views leads
to the multiplicity of functions to serve many purposes.
Consider first the issue of supporting multipurpose through multilevel.
Each level in a multilevel hierarchy represents a particular abstraction and
description. In general, we may use different vocabularies and languages to
describe different levels [80]. In this way, the same thing is described and
represented in multiple ways. As a result, we can ask the right questions at
the right level by using the right languages and for the right purposes.
The point may be illustrated by using software system development as
an example. Normally, a software system is developed in multiple stages in
which a stage corresponds to a particular level of description. At the initial
stage, the requirements and high-level specifications may be given by using a
natural language, possibly aided by figures and tables. We ask the question
of what is the system. The main purpose is to specify the functions of the
system. Once we are satisfied with specifications, we may move to a stage
of designing a system architecture. We ask the question of what are the
components of the system and their functions. The purpose is to provide a
blueprint for building the system. At a final stage, we focus on the imple-
mentation by using a programming language. We ask the question of how
to implement the functions of the various components. The main purpose is
to build an operational system. Although such a three-stage or three-level
description of software system development is an oversimplification of a com-
plex process involving many stages or levels, it is sufficient for the purpose of
illustrating our point. The idea of dividing a complex process into multiple
stages or levels for serving multiple purposes is our main concern, and the
actual number of stages or levels is less crucial and less relevant.
When explaining how language works, Crystal [19] considers a five-level
hierarchy for syntactic investigation, consisting of sentences, clauses, phrases,
words, and morphemes. Rules for constructing the five types of elements of

18
a language are not entirely the same. As a result, it may be necessary to
have different methods for representation, description, and investigation. The
purposes of investigations at different levels are different and we ask different
questions at the different levels. The example again supports the argument
that multilevel leads naturally to multipurpose.
Similarly, we can consider the issue of supporting multipurpose through
multiview. This can be easily done by drawing a correspondence between
levels and views. Different views describe the same thing from different an-
gles and possibly by different languages. A particular view typically makes a
certain aspect become more apparent for a specific purpose and pushes other
aspects into the background. To illustrate this point, we use a multiple hi-
erarchy model of social stratification proposed by Jeffries and Ransford [30].
They consider social inequality hierarchies formed, respectively, by class, eth-
nicity, gender, and age. While each hierarchy serves the purpose of enabling
us to identify one type of social inequality, it fails to identify other types.
Moreover, they argue that a study of social stratification should not focus on
isolated and fragmented views, but holistic and unified views by using multi-
ple hierarchies. In other words, while each hierarchy may serve one purpose,
an integration of many hierarchies may serve multiple purposes. We can eas-
ily find many more examples to support the argument that multiview leads
naturally to multipurpose [14].

4.2. Triarchic theory of granular computing: philosophy, methodology, and

mechanism
A motivation for introducing the triarchic theory is a plea for a holistic
study of granular computing in response to an imbalanced research effort [73].
Existing studies have been focused on machine-oriented and process-based
approaches to granular computing. We concentrate on specific models and
methods by making use of information granules. We pay less attention to the
philosophical foundations and the methodological questions of granular com-
puting. The triarchic theory unifies the three aspects of granular computing:
philosophy, methodology, and mechanism. In this way, granular computing is
both human-oriented and machine-oriented [78]. Results from studies on the
philosophy and methodology aspects empower us with strategies, principles,
methods, and cognitive tools, so that we become better at solving problems.
Results from studies on the mechanism aspect will play a valuable role in the
design and implementation of intelligent systems [73, 78].

19
 Philosophy of structured thinking. Granular computing views the
world through a lens of structures. We represent, interpret, and study a
complex system or a complex problem with the aid of a web of granules
called a granular structure. By taking a structured view of the world, we turn
complexity into simplicity. The philosophy of granular computing is granular
thinking. It is about thinking structurally by making use of granules, levels,
and hierarchies to form useful structures. Granular computing adopts ideas
from other philosophical thinkings, including reductionist thinking, systems
thinking, and levelist thinking. Granular computing also combines analytical
thinking and synthetical thinking.
Methodology of structured problem solving. Granular computing
is about using granular structures in the process of problem solving, in which
granular structures can be either explicitly built or implicitly embodied in the
process. Methodology of granular computing concerns about general princi-
ples and application-independent ways to problem solving. For example, an
ordering of levels according to their granularity suggests several approaches,
such as top-down, bottom-up, and middle-out methods. These approaches
have proved to be effective in computer programming. The underlying princi-
ples can be equally applicable to other types of problem solving, for example,
structured scientific investigation and structured writing.
Mechanism of structured information processing. Both the phi-
losophy and methodology guide us designing structured approaches to in-
formation processing in machines. Studies on mechanism turn attention to
architectures of granular computing for structured information processing. In
other words, we explore concrete ways to implement structured information
processing. To make effective use of granular structures, we may consider
graph-theoretic representations of architectures of granular computing. In a
graphic model, a graph itself gives a topological architecture, in which nodes
and components of a graph represent granules and edges represent relation-
ships among granules.
In many situations, the three aspects are interwoven together and they are
integrated into a comprehensive whole. Their separation is for the purpose
of understanding, in line with a general principle of three-way decision. The
three components of the triarchic theory mutually support each other and
any one of them is indispensable. The triarchic theory studies both the
separation and integration of the three aspects.

20
4.3. Three modes of multilevel computing: top-down, bottom-up, and middle-
out
A hierarchy consists of multiple levels of differing granularity. The order-
ing of levels according to their granularity represents a coarsening-refinement
relation, which may be interpreted in terms of control-support, generalization-
specialization, and abstraction-concretization. According to the ordering, we
have at least three modes of multilevel information processing [73].
Top-down mode. Top-down approaches are analytic thinking and may
be considered as conceptually driven methods of information processing [41].
We work from higher levels of granularity downwards to lower levels of gran-
ularity. As we move towards lower levels, we gradually add more details.
This makes an abstract understanding more concrete. The correctness of a
higher level ensures the correctness of subsequent lower levels. Thus, for a
top-down approach to work, we must have a global view and a conceptual
understanding of the whole problem.
Bottom-up mode. Bottom-up approaches are synthetical thinking and
may be interpreted as data-driven methods [41]. We work from lower levels
of granularity upwards to higher levels of granularity. By extracting the most
common and useful features, we build a higher level abstraction supported
by its lower details. Details at a lower level normally suggest many possible
abstractions at a higher level. For a bottom-up approach to work, we must
construct a best abstraction at each level based on the evidence from its
lower level.
Middle-out mode. Middle-out approaches combine both top-down and
bottom-up thinkings. We start at a level that we have sufficient information
and a good understanding. This level may be viewed as a basic level that
provides goals for guiding top-down investigations and evidence for support-
ing bottom-up explorations. Based on an understanding and goals formed
at the basic level, we move downwards to develop more details. At the same
time, we move upwards for further abstraction.
Top-down approaches require a global understanding for guiding succes-
sive divisions. However, we may not have a global view unless we know some
details of its components. That is, an understanding of a whole requires some
understanding of its parts. On the other hand, bottom-up approaches may
suffer from a lack of a clearly stated overall goal for successive combination.
That is, an understanding of parts requires some understanding of the whole.
Middle-out approaches are based on a partial understanding of the whole and
the parts. For effective problem solving, any one of the three modes may not

21
be sufficient. We may use different modes at different stages of a problem
solving process [2, 35, 61]. In a bottom-up manner, we can form some goals
based on available information. Based on the formed goals, we can have
further top-down investigations. At the same time, we can take middle-out
approaches to help us to attain a better understanding at a particular level.
It is expected that we may use the three modes iteratively and alternatively.

5. Three-way granular thinking

Granular computing is granular thinking at multiple levels of granular-
ity and with multiple granules at each level. By applying the principles of
three-way decision to granular thinking, we have a conception of granular
computing in threes. This specific type of granular computing in threes or
three-way granular computing uses three levels or three granules. Three
granules and three levels may be further interpreted based on other notions
that have a more concrete physical meaning, as explained in the last section.

5.1. The use of threes

Granules may be viewed as the basic units of information and knowl-
edge, vocabulary of discussion, or focal points of investigation. Depending
on particular domains and applications, we have different ways to interpret
granules. For example, granules may represent types, classes, elements, com-
ponents, stages, periods, and many others. To support the use of three levels
and three granules for modelling granular computing, we first collect some
statistical data about the use of different numbers.
Table 1 summarizes the results of Google searches that we did on February
4, 2018 by using exact phrases of the form "n g of", where n takes values
given in the first column of the table and g takes values given in the first
row of the table. Each cell is the number of results reported by Google.
For example, the number reported for a search of the exact phrase "three
levels of" is 11,800,000. Each column of the table, except for the first one,
gives a frequency distribution with respect to numbers from two to seven. For
example, the second column of the table, labelled by “levels,” summarizes the
frequencies of the use of different numbers of levels. Levels of granularity are
a fundamental concept to describe granular structures [80]. In general, levels
may represent layers, extends, or degrees of something. The next six columns
represent three different scenarios: “types” and “classes” for classification and
categorization, “elements” and “components” for three parts of a whole, and

22
 levels types classes elements components stages periods
two 12,300,000 54,700,000 9,550,000 7,090,000 9,570,000 4,970,000 3,320,000
three 11,800,000 27,000,000 4,940,000 4,500,000 5,710,000 6,970,000 1,660,000
four 2,370,000 9,480,000 2,130,000 1,310,000 2,100,000 3,220,000 233,000
five 920,000 3,170,000 493,000 636,000 418,000 596,000 94,500
six 575,000 739,000 264,000 444,000 247,000 388,000 77,600
seven 514,000 456,000 117,000 199,000 66,000 619,000 45,800
Table 1: Statistics on the use of numbers two to seven

“stages” and “periods” for progressive development and evolution processes.

The four representative categories were chosen so that both three ordered
and unordered things are considered, which provides a reasonable coverage
of things in reality. From Table 1, we have the following three observations
and plausible explanations.
Dominance of two, three and four. In all cases, two, three, and
four play dominant roles, while five, six, and seven play limited roles. These
results are consistent with findings regarding the capacity of short-term mem-
ory. For the purpose of cognitive simplicity, it is effective to consider only
a small number of entities or things. In the light of Table 1, the number is
3 ± 1 with 3 as the middle point.
Fast decreasing speed. Except for “stages,” thinking in twos seems to
be the most common theme, followed by thinking in threes and thinking in
fours. Except for “levels” and “stages,” thinking in threes is about half of
thinking in twos, and thinking in fours is less than half of thinking in threes.
There is a large drop when moving from four to five. Let us consider only
the first two levels in the framework given by Warfield [69], that is, given n
things we only count the numbers of individuals and pairs. We have 3+3 = 6
for three, 4 + 6 = 10 for four, and 5 + 10 = 15 for five. They suggest that, if
pair-wise connections are considered, five things may be too many to process
effectively.
“Stages” as an outlier. The number three is the magical number for
“stages.” When it comes to “stages,” a two-stage thinking may be too simple
and cannot provide a good explanation of real phenomena. A three-stage
thinking may bring additional insights. It is also interesting to note that
seven-stage thinking is often used. A plausible explanation is that there
exist some well-accepted three-stage and seven-stage theories and models.

23
For example, a Google search of “three stages of” produces top-ranked topics
of law of three stages, three stages of inflammation, and three stages of
marriage. A Google search of “seven stages of” produces top-ranked topics of
seven stages of grief, seven stages of a man’s life, and seven stages of action.
Although our statistical data is relatively small, it still provides some im-
portant implications. Compared with thinking in twos, thinking in threes is
not as popular and natural. This observation motivates a turn of attention
to thinking in threes in order to avoid problems caused by polarized and du-
alistic thinking [54]. If we think in threes as naturally as we think in twos, we
will be more effective in solving many problems. In some situations, although
thinking in twos is a preferred choice, we are forced to think in threes in order
to cope with uncertainty or a lack of information. When we are confident
and have sufficient information, we choose to think in twos, otherwise, we
think in threes as an approximation. The main idea of sequential three-way
decision is to construct a sequence of approximations in the search of a final
two-way decision [37, 58, 75].
One of the goals of three-way granular computing is to articulate thinking
in three levels and thinking in three granules. Thinking in threes is widely
practiced, which suggests that granular computing in threes is a promising
direction of research.

5.2. Thinking in three levels

Three-level, three-layer, and three-tier structures may provide a simple
and yet comprehensive description of a complicated issue or a complex sys-
tem. Thinking in three levels divides a whole into three levels and explores
a natural ordering of the three levels. At each level, we may use a family
of granules. A granular structure with three levels and multiple granules at
each level is called a 3LmG structure, which is depicted in Figure 3. With a
3LmG structure, it is assumed that granules at each level are relatively inde-
pendent and the three levels are interdependent. Under these assumptions,
a large number of granules at each level may not increase the complexity, as
we can consider the granules individually. The focus of attention will be on
the relationships of levels.
In Figure 3, a top-down reading of the three levels shows a relation of
control and a bottom-up reading shows a relation of support. The top level
and the bottom level are indirectly related by the middle level, rather than
directly related. In this way, we have three individual levels and four tran-
sitions between levels (i.e., two between the top level and the middle level,

24
 top level

control
middle level

support
bottom level

Figure 3: 3LmG structure: Three levels and multiple granules in each level

and another two between the middle level and the bottom level). There are a
total of seven items to be considered (i.e., three levels plus four transitions).
However, in most cases we do not need to consider all seven items at the
same time. In fact, for a particular transition, we only need to consider two
adjacent levels and the transition itself, which is a total of three items. Each
level has its own specific focus and purpose, and may be represented and
studied by means of many granules of a similar nature or granularity. It is
normally required that granules in each level are relatively independent or at
least nearly independent.
One example of 3LmG structures is the three levels of government of
Canada, consisting of the federal level, the provincial level, and the munic-
ipal level. Another example is a three-stratum model of cognitive abilities
proposed by Carroll [12], which consists of the Stratum I of narrow specific
abilities, Stratum II of broad abilities, and Stratum III of general abilities.
We briefly describe three additional three-level models that are related to
computing.
Marr’s three-level understanding of information processing sys-
tems. Marr [46] argues that a full understanding of an information process-
ing system requires understanding at multiple levels. We explore different
kinds of explanations by using different levels of descriptions. Marr suggests
in particular a framework based on three levels:
1. The computational theory level deals with computation in the abstract.
Computation is understood as mappings between different kinds of in-
formation. Their abstract properties, appropriateness, and adequacy
are some of the main concerns.

25
 2. The representation and algorithm level addresses the problem of imple-
menting computation through algorithmic processes, based on a proper
representation of information and associated manipulating processes.
3. The hardware implementation level is about the realization of repre-
sentation and algorithms by using physical devices.
A series of explanations, from abstract to representation and algorithms, and
to physical devices, enables us to have a full understanding of an information
system in its fullest sense.
Three-level database architecture. ANSI-SPARC architecture of
database management systems consists of three levels, namely, the exter-
nal level, conceptual level, and internal level [29]. The three levels address
three different types of questions and issues:
1. The external view describes the data from the user’s perspective, that
is, a user view of the data.
2. The conceptual view describes how the data is represented and pro-
cessed conceptually, as well as how the data is inter-related, that is, a
logic view of the data.
3. The internal level focuses on how the data is physically stored by using
the computer hardware, that is, a physical view of the data.
The separation of the three levels allows data independence so that changes
made at one level do not require extensive changes at other levels.
Weaver’s three levels of communications problems. In presenting
a general framework for communications, Weaver [60] suggests to ask serially
questions at three levels:
1. Level A of the technical problem: How accurately can the symbols of
communication be transmitted?
2. Level B of the semantic problem: How precisely do the transmitted
symbols convey the desired meaning?
3. Level C of the effectiveness problem: How effectively does the received
meaning affect conduct in the desired way?
The three levels concern questions of progressive difficulty. Level A focuses
on the engineering aspects of communication, which is beautifully addressed
by Shannon’s mathematical theory of communication [60], also known as
Shannon’s information theory. The other two levels involve the semantics
and the success in realizing the desired value of the transmitted information,

26
which have not been satisfactorily solved. An insight from Weaver’s three-
level model is dividing a hard problem into simpler problems and solving the
simplest problem first.
As demonstrated by the three examples, there are several advantages
of thinking in multiple levels. First, a totally ordered sequence of levels
provides a natural and simple way for a progressive investigation, from simple
and concrete levels to complex and abstract levels. In a top-down manner,
we can gradually develop and concretize ideas. In a bottom-up manner,
we can abstract essential ideas by removing unnecessary details. Second, a
separation of levels allows us to focus on different types of issues at different
levels. This normally leads to clarification and simplification, so that the
right questions are asked at the right level, as well as avoiding distractions
and confusions from asking the wrong questions. Third, a sequence of levels
offers efficiency in processing. At each level, we typically consider a subset of
all problems and questions, which reduces the complexity. Furthermore, the
ordering of levels reflects the dependencies of problems and questions, which
leads to structured approaches and solutions.
As a special case of multilevel thinking, thinking in three levels avoids
an over-simplification of thinking in two levels and, at the same time, does
not suffer from a complexity of thinking in four or more levels. It is not
surprising to observe examples of thinking in three levels across many differ-
ent disciplines. In philosophy, we study ethics at three levels of meta-ethics,
normative ethics, and applied ethics. In physics, we have three levels of
micro-approaches, meso-approaches, and macro-approaches. In economics,
we have microeconomics, mesoeconomics, and macroeconomics. In biology,
we have three levels of biodiversity, that is, genetic diversity, species diversity,
and ecosystem diversity. In management, we have three levels of low-level
management, middle-level management, and top-level management. These
examples may be sufficient for illustrating the value and effectiveness of think-
ing in three levels and we can easily expand this list of examples.

5.3. Thinking in three granules

Dividing a whole into three components or three parts and looking at the
whole in terms of three dimensions or three perspectives are another type of
common practice and heuristic of human problem solving. We may consider
each component, part, dimension, or perspective as a granule. Furthermore,
we can divide a granule into three smaller granules and obtain a multilevel
granular structure. Such a granular structure with multiple levels and three

27
 1

1 2 3

1 2 3 1 2 3 1 2 3

1 2 3

Figure 4: mL3G structure: Multiple levels and three granules

granules is called an mL3G structure, as depicted in Figure 4. With an mL3G

structure, it is assumed that three granules are interdependent, mutually
support each other, and are individually indispensable. Moreover, granules
at different levels are less dependent and can be investigated in separation.
The focus of attention will be on the relationships of the three granules.
Following the argument of Warfield [69] discussed earlier, we give a frame-
work of three-level analysis for granular thinking in three granules in Figure 5.
The bottom level of individual analysis focuses on one of the three granules
without considering the other two granules. That is, we concentrate on prob-
lems that are locally related to one granule. The middle level of pair-wise
comparative analysis examines one granule in relation to another granule.
This normally involves a comparative study of two granules in a search for
commonalities and differences. The pairs in the middle level correspond to
the notion of orthopairs introduced by Ciucci [15, 16]. Finally, the top in-
tegrative analysis examines three granules based on results from individual
analysis and pair-wise comparison. Through three-level analysis, we may
arrive at a full understanding of the whole.

Integrative analysis (1,2,3)

Pair-wise comparative analysis (1,2) (1,3) (2,3)

Individual analysis (1) (2) (3)

Figure 5: Computing in three granules

28
 The triarchic theory of granular computing, the three elements of gran-
ular structures (i.e., granules, levels, and hierarchies), and the three modes
of computing (i.e., top-down, bottom-up, and middle-out), as discussed in
Section 4, are examples of thinking in threes. We briefly mention three more
examples.
Sternberg’s triarchic theory of intelligence. Sternberg [65, 66] de-
velops a triarchic theory of human intelligence by considering three aspects
of intelligence, namely, creative, analytical, and practical:
1. Creative abilities are related to the creation, invention, and generation
of novel ideas or the design of new things.
2. Analytical abilities are related to analysis, comparison, and evaluation
of ideas, in order to ascertain whether the ideas are good ones.
3. Practical abilities are related to the implementation of ideas, concerning
the realization of the value of ideas through application and utilization.
It is interesting to note that Sternberg [65] describes the triarchic theory
by using a tree-structure similar to Figure 4. He divides the theory into
three subtheories, consisting of an experiential subtheory, a componential
subtheory, and a contextual subtheory. Each subtheory is further divided.
For example, the componential subtheory is divided into a theory of fluid
abilities and a theory of crystallized abilities. The theory of fluid abilities is
again further divided into a theory of induction and a theory of deduction.
The first level of Sternberg’s structure is the same as given by Figure 4.
However, in subsequent levels, Sternberg uses two parts or one part. Using
the terms of granular computing, this is equivalent to saying that a granule
is not necessarily divided into exactly three granules.
Clayton’s three components of influence. Clayton [17] gives three
components of influence that determine the amount of one’s influence:
1. The influencer: We play an essential role in the influencing process. Our
confidence and presence are related to the effectiveness of our influence.
2. The message: How we craft a structured and compelling message is a
prerequisite for achieving influence.
3. The ways of delivering: The way to deliver our message is equally
important, if not more, as the message. We need to be able to negotiate,
to persuade, and to take others’ perspectives.
The trisection of influence is the first, and relatively the simpler, step. Similar
to Parr’s [48] theory of captivology, Clayton’s [17] theory of influence is more

29
about building strategies with respect to the three components of influence.
In other words, the TAO model of three-way decision may also shed some
light on the three-component based theory of influence.
Keidel’s triadic pattern. In an attempt to organize and digest the
vast amount of business literature, Keidel [33] puts forward a triangular
framework by studying organizational strategies with respect to the following
three types of relationships:
1. Disjunction (i.e., non-overlapping) and autonomy: In set-theoretic terms,
Keidel’s notion of disjunction is non-overlapping. That is, two parts
are separated and independent, leading to autonomy of each.
2. Intersection (i.e., overlapping and non-containment) and cooperation:
Keidel’s notion of intersection corresponds to the case in which two
sets have a non-empty overlap and any one is not a subset of the other.
The two parts are related and dependent, but not in a boss/subordinate
relation. In this case, the cooperation of the two parts is needed.
3. Containment and control: The containment may be interpreted as a
boss/subordinate relation, showing a controlling and following relation-
ship.
Organizational strategies and designs can be explained in terms of tradeoff
of three variables corresponding to autonomy, cooperation, and control. In
particular, Keidel uses a triangle to describe mutual supports of the three,
leading to an easy-to-understand and simple-to-use structure for organizing
and understanding business literature and practice.
As demonstrated by these examples, thinking in threes turns complexity
into simplicity. Similar to the case of thinking in three levels, we can also
observe plenty of examples of thinking in three parts across different disci-
plines. In understanding Peirce’s theory of signs, one may use a three-part
basic sign structure consisting of the signifying element of signs, the object,
and the interpretant [3]. Ball [6] considers nature’s patterns as “a trilogy
composed of Shapes, Flow, and Branches.” Kagan [32] looks at the modern
academy according to the three cultures of natural sciences, social sciences,
and the humanities.
One may apply the three-level framework, as given by Figure 5, to sys-
tematically study a whole through the analysis of the three parts and their
interdependencies.

30
6. Concluding remarks
In this paper, we explain three-way decision and three-way granular com-
puting as examples of thinking in threes. The use of threes is predeter-
mined by our limited capacity of information processing, as constrained by
short-term working memory. We present three different types of argument
to demonstrate the needs for and the benefits of thinking in threes. First,
we show mathematically that, under some assumptions, the number three is
the optimal choice for reducing complexity in a search for a cognitive sim-
plicity. The argument is based on the optimality of the base three system.
In one sense, thinking in threes may be viewed as a natural law, although
further investigation is needed for supporting such a strong claim. Second,
we compile everyday uses of numbers by using the Google search engine.
The results show that the number three plays a dominating role. In another
sense, thinking in threes may be viewed as an useful empirical law. As future
work, it is necessary to collect more statistical data in support of thinking in
threes. Third, throughout the paper, we present many examples of thinking
in threes, for example, Marr’s [46] three-level view of information processing
systems, Sternberg’s [65] triarchic theory of human intelligence, Yao’s [73]
triarchic theory of granular computing, Parr’s [48] theory of captivology,
Clayton’s [17] theory of influence, and others. In a third sense, thinking in
threes is a type of practical cognitive tools to deal with the complexity of real
world problems. Further research may be focused on collecting and analyzing
these cognitive tools, in order to have a general theory of three-way decision
as thinking in threes.
In a 2015 white paper entitled “Computing, cognition and the future of
knowing,” Kelly [34] suggests a three-era evolution model of computing: the
tabulating era (1900s-1940s), the programming era (1950s-onward), and the
cognitive era (2011- ). The model by itself is another example of thinking in
threes. To build theory and models of cognitive computing, as a prerequi-
site we must first understand fully human ways to think. To make computer
systems effective, machines should explain their results in human understand-
able terms. Three-way decision, as a special class of human ways to think, is
timely and fits well in the cognitive era of computing [74]. The discussions
of this paper are aimed at drawing a reader’s attention to three-way decision
and three-way granular computing in the cognitive era of computing. The
discussions are therefore presented at a more abstract and conceptual level
by omitting many technical details. The paper provides a long-term and,

31
perhaps personal, vision for future research. The value of thinking in threes
must be fully explored for future human and machine problem solving.
By applying the principles of thinking in threes, we propose a TAO model
of three-way decision. The TAO indicates three main components of three-
way decision, that is, trisecting (T), acting (A), and outcome (O). It is crucial
to combine the right trisection with the right actions to achieve the right
outcome. Existing studies on three-way decision concentrate mainly on the
task of trisecting. It is equally important to investigate the two tasks of
acting and outcome evaluation.
We use the principles of thinking in threes to explain granular computing.
According to the triarchic theory, we investigate granular computing from
three aspects: philosophy of structured thinking, methodology of structured
problem solving, and mechanism of structured information processing [73].
Granular structures result in structuredness in granular computing. We ex-
plain granular structures based on three main ingredients: granules, levels,
and hierarchies. Levels in one hierarchy give a multilevel understanding.
Many hierarchies offer a multiview understanding. The ordering of levels ac-
cording to their granularity suggests three modes of information processing:
top-down, bottom-up, and middle-out approaches. An integration of three
types of approaches is useful for investigating the mechanism of granular
computing.
Three important points that have not been discussed but need comment-
ing.
“Three” as “a few.” In the discussion of three-way decision and three-
way granular computing, we make solely and explicitly use of the magical
number three. We consider the use of three levels and three granules. For
example, in Figure 4 we divide a whole into exactly three granules and fur-
ther divide a granule into exactly three smaller granules. Although there is
overwhelming evidence in support of the use of threes, we should not take
too literal sense of three. Sometimes, we should consider more than three
levels or more than three granules. Other times we may only need to con-
sider two or one. In a more practical sense, we may view “three” as “a few”
when interpreting the principles of three-way decision and three-way granu-
lar computing. Nevertheless, as shown by Table 1, three plays a dominating
role and should be taken as a first choice. The multiple level tree structure
of Sternberg’s [65] triarchic theory of human intelligence may be considered
as an example of thinking in threes, if we interpret three as a few.
Strengthening our cognitive toolbox through “thinking in threes.”

32
A reviewer of the paper states, “while reading the paper it seems sometimes
that we are obliged, whatever the cost, to think in three.” The statement
correctly points out an unintended bias of our arguments and examples in
support of thinking in threes. Table 1 shows that thinking in twos are so
popular and natural. We are more inclined to thinking in twos. The table
also shows that thinking in other numbers is possible. By interpreting “three”
figuratively as “a few,” we correct the bias to a certain degree. In order to
put thinking in threes in its right perspective, we must acknowledge that a)
thinking in threes is only one of the cognitive tools, b) thinking in threes
should not be considered as a denial or replacement of other ways to think,
and c) thinking in threes is complementary to, rather than competitive with,
other ways to think. Our arguments in this paper may be taken as a plea for
attention to thinking in threes. By integrating thinking in threes with other
ways to think, we may be able to strengthen our cognitive toolbox.
“Levels” and “hierarchies” as “granules.” Granules are parts of a
whole. The whole determines the interpretation of parts and parts assign
meaning to the whole. This part-whole setting for studying granules allows
us to use different semantical interpretations of the notion of granules. What
are parts and what is the whole depend on a standpoint. A part may be
viewed as a whole if we focus on the part and look at its internal structure.
A whole may be a part when forming another whole. In our discussion, we
treat granules, levels, and hierarchies as different types of elements. In the
part-whole setting, they may all be treated as different types of granules. For
example, we consider granules as parts when forming a level as the whole of
a family of granules. When considering a hierarchy, levels are parts and a
hierarchy of a family of levels is a whole. In this case, levels may also be
viewed as a different type of granules. When considering multiple hierar-
chies, hierarchies are parts and a family of hierarchies is a whole. That is,
hierarchies may be viewed as another type of granules. The abstract notion
of granules may be attached with different semantics and, consequently, is
labelled differently. Nevertheless, some general principles remain to be the
same, namely, a whole is a family of parts called granules and relationships
of parts give rise to granular structures. This understanding of granules and
granular structures allows the required flexibility of granular computing.

33
Acknowledgements
The author thanks editors and reviewers for their critical and constructive
comments. This work was supported in part by a Discovery Grant from
NSERC, Canada.

References
[1] M.K. Afridi, N. Azam, J.T. Yao, E. Alanazi, A three-way clustering
approach for handling missing data using GTRS. International Journal
of Approximate Reasoning 98 (2018) 11-24.

[2] J.I. Allen, E.A. Fulton, Top-down, bottom-up or middle-out? Avoid-

ing extraneous detail and over-generality in marine ecosystem models.
Progress in Oceanography 84 (2010) 129-133.

[3] A. Atkin, Peirce’s theory of signs. Stanford Encyclopedia of Phi-

losophy, https://plato.stanford.edu/entries/peirce-semiotics/
(accessed March 16, 2018).

[4] R.C. Atkinson, R.M. Shiffrin, Human memory: A proposed system and
its control processes. In: K.W. Spence, J.T. Spence (Eds.), The Psy-
chology of Learning and Motivation, Academic Press, New York, pp.
89-195, 1968.

[5] N. Azam, J.T. Yao, Game-theoretic rough sets for recommender sys-
tems. Knowledge-Based Systems 72 (2014) 96-107.

[6] P. Ball, Nature’s Patterns: A Tapestry in Three Parts, Oxford Univer-

sity Press, Oxford, 2009.

[7] A. Bargiela, W. Pedrycz (Eds.), Human-Centric Information Processing

Through Granular Modelling, Springer, Berlin, 2009.

[8] A. Bargiela, W. Pedrycz, Granular Computing: An Introduction,

Kluwer Academic Publishers, Boston, 2002.

[9] S. Brown, Trinitarianism, the eternal evangel and the three eras of
schema. In: J. Bell, S. Brown, D. Carson (Eds.), Marketing Apocalypse,
Routledge, London, pp. 23-43, 1996.

34
[10] F. Cabitza, D. Ciucci, A. Locoro, Exploiting collective knowledge with
three-way decision theory: Cases from the questionnaire-based research.
International Journal of Approximate Reasoning 83 (2017) 356-370.

[11] F. Capra, The Web of Life, Anchor Books, New York, 1997.

[12] J.B. Carroll, Human Cognitive Abilities: A Survey of Factor-Analytical

Studies, Cambridge University Press, New York, 1993.

[13] S.H. Chen, Y.R. Du, Granularity in economic decision making: An inter-
disciplinary review. In: W. Pedrycz, S.M. Chen (Eds.), Granular Com-
puting and Decision-Making, Springer, Berlin, pp. 47-71, 2015.

[14] Y.H. Chen, Y.Y. Yao, A multiview approach for intelligent data analysis
based on data operators. Information Sciences 178 (2008) 1-20.

[15] D. Ciucci, Orthopairs and granular computing. Granular Computing 1

(2016) 159-170.

[16] D. Ciucci, Orthopairs: A simple and widely used way to model uncer-
tainty. Fundamenta Informaticae 108 (2011) 287-304.

[17] M. Clayton, Brilliant Influence: What the Most Influential People Know,
Do and Say, Prentice Hall, New York, 2011.

[18] N. Cowan, The magical number 4 in short-term memory: A reconsidera-

tion of mental storage capacity. Behavioral and Brain Sciences 24 (2000)
87-185.

[19] D. Crystal, How Language Works, Penguin Books, London, 2006.

[20] D. Dubois, H. Prade, Bridging gaps between several forms of granular

computing. Granular Computing 1 (2016) 115-126

[21] Engineering Research Associates, High-speed Computing Devices,

McGraw-Hill, New York, 1950.

[22] H. Fujita, T.R. Li, Y.Y. Yao, Advances in three-way decisions and gran-
ular computing. Knowledge-Based Systems 91 (2016) 1-3.

[23] C. Gao, Y.Y. Yao, Actionable strategies in three-way decisions.

Knowledge-Based Systems 133 (2017) 141-155.

35
[24] F. Gobet, G. Clarkson, Chunks in expert memory: Evidence for the
magical number four ... or is it two? Memory 12 (2004) 732-747.

[25] B. Hayes, Third base. American Scientists 89 (2001) 490-494.

[26] J.R. Hobbs, Granularity. In: Proceedings of the Ninth International

Joint Conference on Artificial Intelligence, pp. 432-435, 1985.

[27] B.Q. Hu, H. Wong, K.F.C. Yiu, On two novel types of three-way deci-
sions in three-way decision spaces. International Journal of Approximate
Reasoning 82 (2017) 285-306.

[28] C.C. Huang, J.H. Li, C.L. Mei, W.Z. Wu, Three-way concept learning
based on cognitive operators: An information fusion viewpoint. Inter-
national Journal of Approximate Reasoning 83 (2017) 218-242.

[29] D.A. Jardine, The ANSI/SPARC DBMS Model, North-Holland, Ams-

terdam, 1977.

[30] V. Jeffries, H.E. Ransford, Social Stratification: A Multiple Hierarchy

Approach, Allyn and Bacon Inc., Boston, 1980.

[31] C.M. Jiang, Y.Y. Yao, Effectiveness measures in movement-based three-

way decisions. Knowledge-Based Systems, 2018, https://doi.org/10.
1016/j.knosys.2018.07.008

[32] J. Kagan, The Three Cultures: Natural Sciences, Social sciences and
the Humanities in the 21st Century, Cambridge University Press, New
York, 2009.

[33] R.W. Keidel, Strategy made simple: Thinking in threes. Business Hori-
zons 56 (2013) 105-111.

[34] J.K. Kelly, Computing, cognition, and the future of know-

ing. 2015. http://www.research.ibm.com/software/IBMResearch/
multimedia/Computing_Cognition_WhitePaper.pdf (accessed Jan-
uary 8, 2016)

[35] R.A. Kinchla, J.M. Wolfe, The order of visual processing: “Top-down,”
“bottom-up,” or “middle-out.” Perception and Psychophysics 35 (1979)
225-231.

36
[36] E.B. Lease, The number three, mysterious, mystic, magic. Classical
Philology 14 (1919) 56-73.

[37] H.X. Li, L.B. Zhang, X.Z. Zhou, B. Huang, Cost-sensitive sequential
three-way decision modeling using a deep neural network. International
Journal of Approximate Reasoning 85 (2017) 68-78.

[38] J.H. Li, C.C. Huang, J.J. Qi, Y.H. Qian, W.Q. Liu, Three-way cognitive
concept learning via multi-granularity. Information Sciences 378 (2017)
244-263.

[39] X.N. Li, B.Z. Sun, Y.H. She, Generalized matroids based on three-way
decision models. International Journal of Approximate Reasoning 90
(2017) 192-207.

[40] X.N. Li, H.J. Yi, Y.H. She, B.Z. Sun, Generalized three-way decision
models based on subset evaluation. International Journal of Approxi-
mate Reasoning 83 (2017) 142-159.

[41] P.H. Lindsay, D.A. Norman, An Introduction to Psychology, 2nd edition,

Academic Press, New York, 1977.

[42] D. Liu, T.R. Li, D.C. Liang, X. Yang, The temporality and spatiality
of three-way decisions. CAAI Transactions on Intelligent Systems, 2018.
DOI: 10.11992/tis.201804045.

[43] D. Liu, D.C. Liang, Generalized three-way decisions and special three-
way decisions. Journal of Frontiers of Computer Science and Technology
11 (2016) 502-510.

[44] D. Liu, D.C. Liang, C.C. Wang, A novel three-way decision model based
on incomplete information system. Knowledge-Based Systems 91 (2016)
32-45.

[45] X.A. Ma, Y.Y. Yao, Three-way decision perspectives on class-specific

attribute reducts. Information Sciences 450 (2018) 227-245.

[46] D. Marr, Vision, A Computational Investigation into Human Represen-

tation and Processing of Visual Information, W.H. Freeman and Com-
pany, San Francisco, 1982.

37
[47] G.A. Miller, The magical number seven, plus or minus two: Some limits
on our capacity for processing information. Psychological Review 63
(1956) 81-97.

[48] B. Parr, Captivology, the Science of Capturing People’s Attention,

HarperOne, New York, 2015.

[49] Z. Pawlak, Rough Sets, Theoretical Aspects of Reasoning about Data.

Kluwer Academic Publishers, Dordrecht, 1991.

[50] W. Pedrycz, Granular Computing Analysis and Design of Intelligent

Systems. CRC Press, Boca Raton, FL, 2013.

[51] W. Pedrycz, The principle of justifiable granularity and an optimization

of information granularity allocation as fundamentals of granular com-
puting, Journal of Information Processing Systems 7 (2011) 397-412.

[52] W. Pedrycz, S.M. Chen (Eds.), Granular Computing and Intelligent

Systems: Design with Information Granules of Higher Order and Higher
Type. Springer-Verlag, Berlin, 2011.

[53] W. Pedrycz, A. Skowron, V. Kreinovich (Eds.), Handbook of Granular

Computing, John Wiley & Sons Ltd, West Sussex, 2008.

[54] A. Philotunus (Ramsey Dukes), The Good, The Bad, The Funny, The
Mouse That Spins, 1992.

[55] J.E. Phythian, Economical number bases. The Mathematical Gazette

53 (1969) 376-379.

[56] R. Plotnik, H. Kouyoumdjian, Introduction to Psychology, 9th edition,

Wadsworth, Belmont, CA, 2011.

[57] J.J. Qi, T. Qian, L. Wei, The connections between three-way and clas-
sical concept lattices. Knowledge-Based Systems 91 (2016) 143-151.

[58] J. Qian, C.Y. Dang, X.D. Yue, N. Zhang, Attribute reduction for se-
quential three-way decisions under dynamic granulation. International
Journal of Approximate Reasoning 85 (2017) 196-216.

[59] R.S. Ren, L. Wei, The attribute reductions of three-way concept lattices.
Knowledge-Based Systems 99 (2016) 92-102.

38
[60] C.E. Shannon, W. Weaver, The Mathematical Theory of Communica-
tion, The University of Illinois Press, Urbana, 1949.

[61] L.P. Shiu, C.Y. Sin, Top-down, middle-out, and bottom-up processes: A
cognitive perspective of teaching and learning economics. International
Review of Economics Education 5 (2006) 60-72.

[62] H.A. Simon, How big is a chunk? Science 183 (1974) 482-488.

[63] P.K. Singh, Three-way fuzzy concept lattice representation using neutro-
sophic set. International Journal of Machine Learning and Cybernetics
8 (2017) 69-79.

[64] A. Skowron, A. Jankowski, S. Dutta, Interactive granular computing.

Granular Computing 1 (2016) 95-113.

[65] R.J. Sternberg, Beyond IQ: A Triarchic Theory of Intelligence, Cam-

bridge University Press, Cambridge, 1985.

[66] R.J. Sternberg, K. Sternberg, J. Mio, Cognitive Psychology, 6th edition,

Wadsworth, Belmont, 2012.

[67] B.Z. Sun, W.M. Ma, B.J. Li, X.N. Li, Three-way decisions approach
to multiple attribute group decision making with linguistic information-
based decision-theoretic rough fuzzy set. International Journal of Ap-
proximate Reasoning 93 (2018) 424-442.

[68] G.Y. Wang, DGCC: Data-driven granular cognitive computing. Granu-

lar Computing 2 (2017) 343-355.

[69] J.N. Warfield, The magical number three – plus or minus zero. Cyber-
netics and Systems 19 (1988) 339-358.

[70] J.T. Yao, N. Azam, Web-based medical decision support systems for
three-way medical decision making with game-theoretic rough sets.
IEEE Transactions on Fuzzy Systems 23 (2015) 3-15.

[71] J.T. Yao, A.V. Vasilakos, W. Pedrycz, Granular computing: Perspec-

tives and challenges. IEEE Transactions on Cybernetics 43 (2013) 1977-
1989.

39
[72] Y.Y. Yao, Interval sets and three-way concept analysis in incomplete
contexts. International Journal of Machine Learning and Cybernetics 8
(2017) 3-20.

[73] Y.Y. Yao, A triarchic theory of granular computing. Granular Comput-

ing 1 (2016) 145-157.

[74] Y.Y. Yao, Three-way decisions and cognitive computing. Cognitive

Computation 8 (2016) 543-554.

[75] Y.Y. Yao, Granular computing and sequential three-way decisions. In:
RSKT 2013. LNCS (LNAI), vol. 8171, pp. 16-27, 2013.

[76] Y.Y. Yao, An outline of a theory of three-way decisions. In: RSCTC

2012, LNCS (LNAI), vol. 7413, pp. 1-17, 2012.

[77] Y.Y. Yao, Artificial intelligence perspectives on granular computing. In:

W. Pedrycz and S.M. Chen (Eds.), Granular Computing and Intelligent
Systems Design with Information Granules of Higher Order and Higher
Type, Springer, Berlin, pp. 17-34, 2011.

[78] Y.Y. Yao, Human-inspired granular computing. In: J.T. Yao (Ed.),
Novel Developments in Granular Computing: Applications for Ad-
vanced Human Reasoning and Soft Computation, IGI Global, Hershey,
Philadelphia, pp. 1-15, 2010.

[79] Y.Y. Yao, Three-way decisions with probabilistic rough sets. Information
Sciences 180 (2010) 341-353.

[80] Y.Y. Yao, Integrative levels of granularity. In: A. Bargiela, W. Pedrycz

(Eds.), Human-Centric Information Processing Through Granular Mod-
elling, Springer, Berlin, pp. 31-47, 2009.

[81] Y.Y. Yao, A partition model of granular computing. LNCS Transactions

on Rough Sets I (2004) 232-253.

[82] H. Yu, C. Zhang, G.Y. Wang, A tree-based incremental overlapping clus-

tering method using the three-way decision theory. Knowledge-Based
Systems 91 (2016) 189-203.

40
[83] X.D. Yue, Y.F. Chen, D.Q. Miao, J. Qian, Tri-partition neighborhood
covering reduction for robust classification. International Journal of Ap-
proximate Reasoning 83 (2017) 371-384.

[84] L.A. Zadeh, Towards a theory of fuzzy information granulation and its
centrality in human reasoning and fuzzy logic. Fuzzy Sets and Systems
19 (1997) 111-127.

[85] H.R. Zhang, F. Min, Three-way recommender systems based on random

forests. Knowledge-Based Systems 91 (2016) 275-286.

[86] X.Y. Zhang, D.Q. Miao, Three-way attribute reducts. International

Journal of Approximate Reasoning 88 (2017) 401-434.

[87] Y. Zhang, J.T. Yao, Gini objective functions for three-way classifica-
tions. International Journal of Approximate Reasoning 81 (2017) 103-
114.

[88] X.R. Zhao, B.Q. Hu, Fuzzy probabilistic rough sets and their corre-
sponding three-way decisions. Knowledge-Based Systems 91 (2016) 126-
142.

41