Granular Computing

https://doi.org/10.1007/s41066-020-00211-9 (0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL PAPER

Set-theoretic models of three-way decision

Yiyu Yao1

Received: 28 November 2019 / Accepted: 21 January 2020

 Springer Nature Switzerland AG 2020

Abstract
The theory of three-way decision is about a philosophy of thinking in threes, a methodology of working with threes, and a
mechanism of processing in threes. We approach a whole through three parts, in terms of three units, or from three
perspectives. A trisecting–acting–outcome (TAO) model of three-way decision involves trisecting a whole into three parts
and acting on the three parts, in order to produce an optimal outcome. In this paper, we further explore the TAO model in a
set-theoretic setting and make three new contributions. The first contribution is an examination of three-way decision with
nonstandard sets for representing concepts under the two kinds of objective/ontic and subjective/epistemic uncertainty. The
second contribution is an introduction of an evaluation-based framework of three-way decision. We present a classification
of trisections and investigate the notion of an evaluation space. The third contribution is, within the proposed framework, a
systematical study of three-way decision with rough sets, interval sets, fuzzy sets, shadowed sets, rough fuzzy sets, interval
fuzzy sets (or equivalently, vague sets, interval-valued fuzzy sets, intuitionistic fuzzy sets), and soft sets.

Keywords Three-way decision  Interval set  Rough set  Fuzzy set  Shadowed set  Rough fuzzy set  Interval fuzzy set 
Soft set  Intuitionistic fuzzy set  Vague set  Interval-valued fuzzy set

1 Introduction three levels, three layers, three generations, three periods,

three stages, three steps, triangles, triads, triplets, and many
The philosophy of three-way decision (3WD), or more others (Yao 2019b). Such a variety of interpretations of
generally three-way computing, is thinking in threes (Yao threes makes three-way decision a useful and universally
2016, 2018, 2019b, 2020). Three-way decision is built on a applicable theory and a practically feasible computational
sound cognitive basis. Our limited capacity of short-term paradigm. In fact, there is a fast-growing interest in theory
working memory necessitates a practical strategy of pro- and applications of three-way decision, as evidenced from,
cessing a few units of information simultaneously, where for example, a number of recent papers published in this
three is a pivot number (Yao 2016). Thinking in threes is a Granular Computing journal (Agbodah 2019; Cai et al.
common practice in human problem-solving in many dis- 2017; Li and Huang 2019; Ma et al. 2019; Mandal and
ciplines and is deeply rooted in many cultures (Dundes Ranadive 2019; Zhang et al. 2019a, b).
1968). By dividing a whole into three relatively indepen- A general TAO (trisecting–acting–outcome) conceptu-
dent and related parts, we turn the complexity of process- alization of three-way decision involves three basic issues:
ing the complex whole into the simplicity of processing the a) dividing a whole into three relatively independent parts,
three parts, which reduces the cognitive and information b) devising strategies to process the three parts, and c)
overload. Modes of thinking in threes include thinking in optimizing trisecting and acting to achieve an expected
three parts, three elements, three components, three cate- outcome. In order to fully realize the power of three-way
gories, three perspectives, three views, three dimensions, decision, it is necessary to investigate formally some
specific models within the TAO framework. For this pur-
pose, the main objective of the present paper is to review
& Yiyu Yao and to examine set-theoretic models of three-way decision
[email protected] in the context of concept understanding, representation,
1
Department of Computer Science, University of Regina, and approximation. An important step is to search for
Regina, SK S4S 0A2, Canada concrete interpretations of various generalized nonstandard

123
 Granular Computing

sets. We examine two specific interpretations. One inter- concept applies. The extension of a concept is the set of
pretation views nonstandard sets as representations of objects that are instances of the concept. In the ideal situ-
concepts having a gradually changing boundary. It captures ation, a concept has a sharp and clear boundary with
the objective/ontic uncertainty, that is, the natural intrinsic respect to its extension and can be precisely defined and
uncertainty, of some concepts. The other interpretation characterized with respect to its intension. In this case, set
views nonstandard sets as representations of concepts with theory and logic are suitable tools. In non-ideal situations, a
a partially known boundary or an indefinable boundary. It concept may have a gradually changing boundary and/or
captures the subjective/epistemic uncertainty concerning our understanding or available information may be insuf-
our understanding of some concepts, due to insufficient ficient to define a concept precisely with respect to either
information and knowledge. Although it is possible to the intension or the extension. This calls for nonstandard
examine other interpretations, the two interpretations are set theories and non-classical logics for dealing with
sufficient for the present study. associated uncertainty.
Set-theoretic models of three-way decision focus on Uncertainty is a multifaceted notion. There are many
decision-making under uncertainty. In situations with types of uncertainty, many sources for producing uncer-
uncertainty, practical decision-making often requires tainty, and many causes of uncertainty. In this paper, we
knowing whether or not an object is an instance of a use various generalizations of sets to represent and char-
concept. The theory of three-way decision suggests a sen- acterize a few commonly studied types of concepts under
sible useful solution: accept or reject, beyond a reasonable uncertainty. We consider the two kinds of ontic and epis-
doubt, only some objects to be members of a concept, and temic uncertainty (Couso and Dubois 2014; Walker et al.
neither accept nor reject the rest of objects. In other words, 2003). The objective/ontic uncertainty is about the inherent
we have a three-way approximation of a concept, consist- characteristics of a concept, and the subjective/epistemic
ing of the sets of instances, non-instances, and undecidable uncertainty is about our understanding of a concept. Fur-
objects. The results will have significant implications to thermore, we restrict the discussions to the following types
data analysis, machine learning, human understanding, and of uncertainty. For objective uncertainty, if a concept has a
intelligent systems, in which the notion of concepts plays sharp/clear boundary, the concept is certain and without
an essential role. uncertainty. If a concept has a gradually changing bound-
The rest of the paper is organized as follows: Section 2 ary, the concept is uncertain. For subjective uncertainty, we
gives a summary of six types of generalized nonstandard look at the uncertainty with respect to intension and
sets for representing concepts under uncertainty. We extension of concepts. In the case of intension, although we
introduce the notions of acceptance, rejection, and non- have precise characterization of a concept in terms of its
commitment rules for producing a three-way approxima- extension, we may not have a precise definition of the
tion of a concept represented by a generalized set. Sec- concept in terms of its intension. This is modeled by the
tion 3 is a review of basic concepts and issues of three-way notion of definability/indefinability of a concept. In the
decision. We present a general TAO (trisecting–acting– case of extension, our understanding or available infor-
outcome) framework of three-way decision and discuss mation is insufficient to specify precisely the extension of a
evaluation-based models for constructing a trisection. Sec- concept, although we may be able to describe qualitatively
tion 4, based on the evaluation-based models, systemati- and linguistically the intension of the concept. This is
cally studies three-way decision with the six types of modeled by the notion of partial knowledge about
nonstandard sets. Section 5 gives some concluding remarks concepts.
and points out future research topics. Table 1 summarizes the uses of sets and six general-
izations of sets for representing the types of uncertainty
mentioned earlier. Rough sets focus on approximations of
2 Nonstandard sets for representing indefinable sets/concepts by using definable sets/concepts
concepts under uncertainty and their (Pawlak 1982, 1991; Yao 2015). Interval sets deal with
three-way approximations partially known sets/concepts (Yao 1993, 2017). Fuzzy sets
model sets/concepts with gradually changing boundaries
One of the fundamental ideas of Port-Royal Logic (Ar- (Dubois and Prade 2012; Zadeh 1965). Rough fuzzy sets
nauld and Nicole 1996) is the understanding and repre- are a generalization of rough sets for approximating fuzzy
sentation of a concept in terms of a pair of an intension (or sets through definable sets (Dubois and Prade 1990).
comprehension) and an extension (or denotation) of the Interval fuzzy sets, also known as mathematically equiva-
concept. The intension of a concept may be viewed as all lent interval-valued fuzzy sets (Dubois 2010; Dubois and
properties, attributes, or more generally some formulas of a Prade 2012), intuitionistic fuzzy set (Atanassov
language that are valid for all those objects to which the 1986, 1999), and vague sets (Bustince and Burillo 1996;

123
Granular Computing

Table 1 Generalized sets for

Types of sets Objective uncertainty Subjective uncertainty
representing different types of
uncertainty Sets Sharp/clear boundary None
Rough sets Sharp/clear boundary Indefinability
Interval sets Sharp/clear boundary Partial knowledge
Fuzzy sets Gradually changing boundary None
Rough fuzzy sets Gradually changing boundary Indefinability
Interval fuzzy sets Gradually changing boundary Partial knowledge
Soft sets Multiply-defined boundary Partial knowledge

Gau and Buehrer 1993), are counterparts of interval sets for applies, with a neither high nor low degree, to the object
modeling partially known fuzzy sets. Soft sets model and/or we have insufficient information or do not have a
concepts either with multiply-defined boundaries or char- strong belief about whether or not the object is an
acterized by partial knowledge (Molodtsov 1999; Yang and instance.
Yao 2020). Three important clarifications of Table 1 need The use of the and/or logic connective enables us to con-
to be made. First, an interpretation of a particular type of sider ontic uncertainty, epistemic uncertainty, or both
generalized sets is only one of many possible interpreta- uncertainties. In other words, yes/no decisions are made
tions. Other interpretations are possible and may be supe- beyond a reasonable doubt. Whenever we are in reasonable
rior to the given interpretation for some applications. doubt and cannot make a yes/no decision for an object
Second, the types of generalized sets in the table are only confidently, we choose a non-commitment decision. The
examples. There are other types of generalized sets. The six result is a three-way approximation of a concept, consisting
types in the table are sufficient for us to demonstrate the of the sets of instances, non-instances, and undecidable
values of three-way decision in a set-theoretic setting. objects, respectively.
Third, there are other types of uncertainty. In summary, the
table may be viewed as a list of examples, rather than an
exhaustive summary. 3 An introduction to three-way decision
Although some concepts may have unsharp and gradu-
ally changing boundaries, our understanding of these con- This section presents a TAO (trisecting–acting–outcome)
cepts is partial, or the available information about these framework of three-way decision. We investigate seven
concepts is incomplete, there is still a practical need to tell classes of trisections and two evaluation-based models for
roughly or qualitatively which objects are instances of a constructing trisections.
concept and which objects are non-instances. This is where
the theory of three-way decision comes into play and 3.1 The TAO of three-way decision
shows its advantages. Instead of making an instance/non-
instance decision for every object, we only make such yes/ Three-way decision is a philosophy of thinking in threes, a
no decisions for some objects and leave the rest of objects methodology of working with threes, and a mechanism of
undecided. More specifically, we have three types of rules processing through threes (i.e., triplets of three elements or
known as the acceptance, rejection, and non-commitment sets of three items). A TAO (trisecting–acting–outcome)
rules: framework of three-way decision is given in Fig. 1, which
• Acceptance rule we accept an object to be an instance focuses on three related tasks (Yao 2018): (1) to divide the
of a concept if the concept applies, with a reasonably whole into three parts, (2) to devise strategies to process
high degree, to the object and/or we have sufficient the three parts, and (3) to optimize a desirable outcome.
information or a strong belief about the object for being The TAO of three-way decision by itself is an example of
an instance; thinking in threes, namely trisecting, acting, and outcome
• Rejection rule we reject an object to be an instance of a optimization. Trisecting divides the whole into three parts,
concept if the concept applies, with a reasonably low acting applies a set of strategies to process the three parts,
degree, to the object and/or we have sufficient infor- and outcome is a desirable combined result of trisecting
mation or a strong belief about the object for not being and action.
an instance; A trisection of the whole is a fundamental notion of
• Non-commitment rule we neither accept nor reject an three-way decision. Each part represents a particular
object to be an instance of a concept if the concept aspect, perspective, or component of the whole; their
integration represents and covers the whole. Each part is

123
 Granular Computing

Definition 1 An un-ordered triplet hX; Y; Zi of subsets X,

Whole Y, and Z of a universe U is called,
(a) a trisection if it satisfies (i),
Trisecting
(b) a non-empty trisection if it satisfies (i) and (ii),
B (c) a distinctive trisection if it satisfies (i) and (iii),
(d) a disjoint trisection if it satisfies (i) and (iv),
A C
(e) a distinctive and disjoint trisection if it satisfies (i),
(iii), and (iv),
Acting (f) a tri-covering if it satisfies (i), (ii), and (iii),
(g) a tri-partition if it satisfies (i), (ii), and (iv).
Strategies Figure 2 is the Hasse diagram of the seven classes of
trisections. Property (i) characterizes the most general class
Outcome of trisections. Tri-coverings and tri-partitions are two
commonly used classes of trisections. Tri-partitions are the
Fig. 1 The TAO of three-way decision (adapted from Yao (2019b)) most specific class of trisections. Many studies of three-
way decision use the class of disjoint trisections (Yao
distinct from and nearly independent of the other two. At 2012). The introduction of seven classes of trisections may
the same time, the three parts are connected. In Fig. 1, the expand the applications of three-way decision.
triplet (A, B, C) forms the triangle of three-way decision, We can broadly divide the seven classes into two
indicating that the three parts are both separated and con- groups, as represented by, respectively, normal rectangles
nected. The combination of trisecting and acting forms the and thick-lined rectangles in Fig. 2. One group consists of
triangular bipyramid with base triangle (A, B, C) of three- the four classes inclusively between trisection and tri-
way decision for the purpose of producing a desired covering in Fig. 2. A trisection satisfying only property (i)
outcome. consists of three subsets whose union covers the entire
universe U. It is possible that one or two of the three
3.2 Classification of trisections subsets are the empty set. It is also possible that two subsets
are the same or all three subsets are the universe (i.e.,
In a set-theoretical setting, we may explain the notion of a hU; U; Ui). For a non-empty trisection defined by (i) and
trisection as follows. The whole is a universal set and a part (ii), all three subsets are not the empty set and some or all
is a subset. A trisection is a triplet of three subsets such that them may be the same. For a distinctive trisection defined
their union is the universal set. One or two subsets may be by (i) and (iii), all three subsets are different and one of
the empty set. In general, it may also be meaningful to them may be the empty set. A tri-covering defined by (i),
allow some overlap between subsets. (ii), and (iii) consists of three different non-empty subsets.
To precisely define different classes of trisections, we In this group of four classes of trisections, any of the three
consider the following list of properties:

Trisection (i)

(i) X [ Y [ Z ¼ U;
(ii) X 6¼ ;; Y 6¼ ;; Z 6¼ ;;
(iii) X 6¼ Y; X 6¼ Z; Y 6¼ Z;
(iv) X \ Y ¼ ;; X \ Z ¼ ;; Y \ Z ¼ ;: Nonempty trisection Distinctive trisection Disjoint trisection
(i), (ii) (i), (iii) (i), (iv)

Property (i) requires that the union of the three subsets

covers the universal set U. By property (ii), all three
subsets are not empty. By property (iii), the three subsets
Tri-covering Distinctive & disjoint
are different from each other. Property (iv) states that the (i), (ii), (iii) trisection (i), (iii), (iv)

three subsets are pairwise disjoint. Property (i) may be

viewed as a basic property of a trisection; any trisection
must have this property. Properties (ii) and (iv) together
imply property (iii). By combining all non-equivalent Tri-partition
(i), (ii), (iv)
subsets of the set of properties (ii)–(iv) with (i), we can
define seven classes of trisections. Fig. 2 Hasse diagram of classes of trisections

123
Granular Computing

subsets may not be defined by using the other two. evaluation. The evaluation-based models reflect a common
Therefore, we must give all three subsets. practice in human judgments, reasoning, and decision-
The other group consists of the three classes inclusively making in terms of bipolarity. Following the discussions of
between disjoint trisection and tri-partition in Fig. 2. A human evaluative processes by Cacioppo et al. (1997),
disjoint trisection defined by (i) and (iii) is composed of Dubois and Prade (2008) and Osgood et al. (1957), we may
three pairwise disjoint subsets. One or two of them may be evaluate objects in U with reference to two landmarks,
the empty set. Non-empty subsets in a disjoint trisection representing positivity and negativity. In other words,
are different. By properties (i) and (iii), we have the fol- evaluations are related to the positive and negative sides
lowing relationships between three subsets (Béziau 2012): and aspects of objects or positive and negative information
about objects. We may view two evaluations as repre-
senting two independent positive and negative evaluative
(a) X c ¼ Y [ Z; Y c ¼ X [ Z; Z c ¼ X [ Y; processes according to two opposite landmark values, such
(b) X ¼ Y c \ Z c; Y ¼ Xc \ Z c; Z ¼ Xc \ Y c; as true–false, good–bad, hospitable–hostile, and many
(c) X  Y c; X  Z c; Y  Xc; Y  Z c; Z  Xc; Z  Y c: other pairs. We may view one evaluation as representing an
integrated evaluative process that considers simultaneously
A distinctive and disjoint trisection defined by (i), (iii), and both positivity and negativity.
(iv) consists of three different and non-overlap subsets, and For developing a model with two evaluations, we use a
one of them may be the empty set. Finally, a tri-partition pair of totally ordered sets ððLþ ; Þ; ðL ; ÞÞ as the scales
consists of three non-empty disjoint subsets. for positive and negative evaluations. We use the same
According to properties (a)–(c), for the second group of symbol  for the two orders in Lþ and L . Like the less-
three classes of trisections, it is sufficient to give only two than relation \ on real numbers,  consistently represents
subsets from hX; Y; Zi and to define the third by using (b). orderings from more negative to less negative and from
With two subsets, we can have two additional representa- less positive to more positive. For two values a; b 2 L , if
tions of a disjoint trisection. Ciucci (2011) introduced the a  b, we also write b  a and say that a is more nega-
notion of an orthopair that is an un-ordered pair of subsets tive than b or b is less negative than a. When a  b or
hX; Yi with X \ Y ¼ ;. Given a disjoint trisection hX; Y; Zi, a ¼ b, we write a4b or b<a. For two values c; d 2 Lþ , if
we have three orthopairs hX; Yi, hX; Zi, and hY; Zi. Con- c  d, we also write d  c and say that c is less positive
versely, given any of the three orthopairs, for example, than d or d is more positive than c. When c  d or c ¼ d,
hX; Yi, we can construct the disjoint trisection we write c4d or d<c.
hX; Y; Zi ¼ hX; Y; ðX [ YÞc i. Yao (1993, 2017) introduced For building a model with one evaluation, we use a
the notion of an interval set as a pair of ordered nested totally ordered set ðL; Þ as the scale of the evaluation. For
subsets ½X; X with X  X. An interval set can be used to two values a; b 2 L, if a  b, we also write b  a and say
define an orthopair hX; X  Xi. Conversely, given an that a is more negative than b, a is less positive than b, b is
orthopair hX; Yi, we can define two interval sets ½X; X [ Y less negative than a, or b is more positive than a. The
and ½Y; Y [ X. For a disjoint trisection hX; Y; Zi, we may interpretations of 4 and  in the positive and negative
construct six possibly different interval sets ½X; X [ Y, evaluations are therefore consistent with the interpretation
½X; X [ Z, ½Y; Y [ X, ½Y; Y [ Z, ½Z; Z [ X, and ½Z; Z [ Y. of  in the single evaluation.
Given any one of the six interval sets, for example, In general, positive and negative evaluations are inde-
½X; X [ Y, we can obtain the disjoint trisection pendent (Cacioppo and Berntson 1994). While the positive
hX; Y; Zi ¼ hX; X [ Y  X; ðX [ YÞc i. In summary, we evaluation uses a scale ðLþ ; Þ, the negative evaluation
have three ways to represent a disjoint trisection: (a) three uses another scale ðL ; Þ. The two evaluative processes
pairwise disjoint subsets, (b) an orthopair, and (c) an are done separately. The result is a pair of values repre-
interval set. Distinct representations provide different senting both the positivity and the negativity of an object.
semantics interpretations, and we may use any one of them In many situations, instead of using two evaluations, it
that is suitable for a particular application. might be simpler and more convenient to have one evalu-
ation by combining the negative and positive aspects of
3.3 Evaluation-based method for constructing objects. In this case, an evaluation is given by using a
trisections totally ordered set ðL; Þ as the scale. It is possible to
combine the two scales ðLþ ; Þ and ðL ; Þ into one
Yao (2012) introduced two models for constructing a tri- bipolar scale (Dubois and Prade 2008). Assume that Lþ and
section based on evaluations of objects. One model uses a L have different values. By taking all values from the two,
pair of evaluations, and the other model uses a single we have a bipolar scale ðL [ Lþ ; fða; bÞ j a; b 2 L ^ a 

123
 Granular Computing

bg [ fða; bÞ j a; b 2 Lþ ^ a  bg [ fða; bÞ j a 2 L ^ b 2 information (Lang et al. 2020; Yao 2019a). The signs
Lþ gÞ. Unlike the case of two evaluations, it is no longer reflect positivity and negativity of evaluations, and the
possible to evaluate an object to be some degree of nega- absolute values reflect the strength of evaluations. In the
tive and, at the same time, another degree of positive. interval ½0; þ1, 0 and þ1 indicate, respectively, the min-
Depending on the number of evaluations, we have two imum and maximum positivity. Similarly, in the interval
models for constructing a trisection of a set of objects (Yao ½1; 0, 1 and 0 indicate, respectively, the maximum and
2012). minimum negativity. Sometimes, 0 is interpreted as rep-
resenting neutrality. Instead of using ½1; 0 for evaluating
Definition 2 Suppose that ep : U ! ðLþ ; Þ is a positive
negativity, one may use [0, 1]. In this case, a value in [0, 1]
evaluation and en : U ! ðL ; Þ is a negative valuation
indicates the degree of negative. We can change [0, 1] to
on a finite non-empty set of objects U. For x 2 U, ep ðxÞ and
½1; 0 by changing a 2 ½0; 1 to a 2 ½1; 0.
en ðxÞ indicate the degree of positivity and the degree of Figure 3 describes the space of a pair of evaluations
negativity of x. Given a pair of thresholds htp ; tn i, tp 2 Lþ with scales ([0,+1],\) and ð½1; 0; \Þ, where the first
and tn 2 L , we divide U into the negative region number in a pair ðvp ; vn Þ represents the value of positive
NEGhtp ;tn i ðep ; en Þ, the positive region POShtp ;tn i ðep ; en Þ, and evaluation and the second number represents the value of
the boundary region BNDhtp ;tn i ðep ; en Þ as follows: negative evaluation. Given a pair of thresholds, we divide
  the evaluation space into four parts. The four parts in turn
NEGhtp ;tn i ðep ; en Þ ¼ x 2 U j en ðxÞ 4 tn ^ ep ðxÞ  tp ;
  are used to define the negative, positive, and boundary
POShtp ;tn i ðep ; en Þ ¼ x 2 U j ep ðxÞ < tp ^ en ðxÞ  tn ; regions in U, as shown in Fig. 3. The three regions are
BNDhtp ;tn i ðep ; en Þ ¼ ðNEGhtp ;tn i ðep ; en Þ [ POShtp ;tn i ðep ; en ÞÞc pairwise disjoint and their union is the universe U, that is,
 
¼ x 2 U j en ðxÞ 4 tn ^ ep ðxÞ < tp [ the triplet hNEGhtp ;tn i ðep ; en Þ; BNDhtp ;tn i ðep ; en Þ; POShtp ;tn i
  ðep ; en Þi is a disjoint trisection of U. The negative region
y 2 U j en ðyÞ  tn ^ ep ðyÞ  tp :
consists of those objects that are evaluated to be highly
The triplet hNEGhtp ;tn i ðep ; en Þ; BNDhtp ;tn i ðep ; en Þ; POShtp ;tn i negative and not highly positive, i.e., en ðxÞ  tn ^ ep
ðep ; en Þi is a disjoint trisection of U. ðxÞ\tp . In contrast, the positive region consists of those
objects that are evaluated to be highly positive and not
Definition 3 Suppose that e : U ! ðL; Þ is an evalua- highly negative, i.e., ep ðxÞ tp ^ en ðxÞ [ tn . The two
tion on a finite non-empty set of objects U. For x 2 U, e(x) regions therefore truly reflect the status of an object for
is evaluation value of x. Given a pair of low and high being either positive or negative as given consistently by
thresholds (l, h), l; h 2 L, with l  h, we trisect U into three both evaluations. The boundary region can be expressed as
regions as follows: the union of two disjoint sets fx 2 U j ep ðxÞ tp ^
NEGð;l ðeÞ ¼ f x 2 U j eðxÞ4lg; en ðxÞ  tn g and fy 2 U j ep ðyÞ\tp ^ en ðyÞ [ tn g. The first
set consists of those objects that are evaluated to be both
BNDðl;hÞ ðeÞ ¼ f x 2 U j l  eðxÞ  hg; ð1Þ
highly positive and highly negative. This situation may
POS½h;Þ ðeÞ ¼ f x 2 U j eðxÞ<hg: happen when we use a pair of non-reciprocal evaluations. It
The disjoint trisection hNEGð;l ðeÞ; BNDðl;hÞ ðeÞ; POS½h;Þ ðeÞi
consists of the sets of negative, neutral, and positive objects.
(+1, −1)
These two models serve as a basis for building set-the-
oretic models of three-way decisions. It needs to be pointed
out that choosing meaningful and easy-to-understand scales BND tn
is one of the fundamental issues. tp

3.4 Evaluation-based models under numerical

(0, −1) NEG POS (+1, 0)
scales
ne

ep

To gain more understanding of the two abstract models, we

ga

BND
ity
tiv

consider examples of scales in terms of numerical values.

iv
ity

sit

For example, we use ð½0; þ1; \Þ and ð½1; 0; \Þ for

po
en

representing, respectively, the values of positive and neg-

ative evaluations. In this way, we are able to capture (0, 0)
explicitly both qualitative information and quantitative
Fig. 3 Trisection based a pair of evaluations

123
Granular Computing

represents a kind of co-existence of positivity and nega- values fall in the center part of the square. These obser-
tivity. The second set consists of those objects that are vations are not surprising, as the center part is where we
evaluated to be both not highly positive and not highly have most uncertainty given by the negative and positive
negative. Those objects are somewhere in the middle evaluations.
according to both positive and negative evaluations. The In some situations, an evaluation function may not
name boundary region seems to reflect better the meaning necessarily be interpreted as representing bipolarity of
of the second set of objects. negativity and positivity. The values represent an ordering
Given a pair of evaluations en : U ! ½1; 0 and or a preference. We may use an evaluation e : U ! ½0; 1
ep : U ! ½0; þ1, we may combined them into an aggre- with a scale [0, 1], where 0 indicates the minimum value
gated evaluation e : U ! ½1; þ1. One possible aggre- and 1 indicates the maximum value. Similar to Defini-
gated evaluation is given by: tion 3, we can trisect the set of objects based on a pair of
thresholds on e.
e ¼ ep þ en : ð2Þ
Definition 4 Suppose that e : U ! ½vmin ; vmax  is an
For any object x 2 U, 1  eðxÞ  þ 1. Under one eval-
evaluation on a finite non-empty set of objects U, where
uation, the value of an object is negative, zero, or positive.
vmin \vmax are two real numbers. Given a pair of low and
For the aggregated evaluation in Eq. (2), objects of similar
high thresholds (l, h) with vmin  l\h  vmax , we trisect U
absolute degrees of negative and positive are evaluated to
into three regions as follows:
be neutral, independent of the actual degrees of negativity
and positivity. A trisection induced by an aggregated L½vmin ;l ðeÞ ¼ f x 2 U j eðxÞ  lg;
evaluation is given in Fig. 4. The dotted line represents all M ðl;hÞ ðeÞ ¼ f x 2 U j l\eðxÞ\hg; ð3Þ
pairs of values such that the aggregated value is 0, that is, ½h;vmax 
en ðxÞ þ ep ðxÞ ¼ 0. As we move right toward ðþ1; 0Þ, the H ðeÞ ¼ f x 2 U j eðxÞ hg:
aggregated values become more positive, and as we move In this case, we can use the range of values as the super-
left toward ð0; 1Þ, the aggregated values become more scripts of the three regions. The disjoint trisection
negative. The thresholds tp and tn are depicted by the two hL½vmin ;l ðeÞ; M ðl;hÞ ðeÞ; H ½h;vmax  ðeÞi consists of the sets of
dashed lines. They divide the evaluation space into three objects with low, medium, and high values.
sections, which leads to a trisection of the set of objects U
defined by Definition 3. The simple aggregated evaluation The low and high values are two opposites and medium
e ¼ ep þ en is only an example for illustration. In general, values are the middle. They may be interpreted as three
it is possible to consider other aggregations of positive and levels, which leads to tri-level thinking (Yao 2019b). There
negative evaluations (Cacioppo et al. 1997). are many examples of such an evaluation-based model. For
The aggregated evaluation does not take consideration example, in medicine it is a common practice to divide
of all information given by a pair of negative and positive some test results into high, normal, and low three levels. In
evaluations. As shown in Figs. 3 and 4, there are some financial management, according to family income, we
degree of consistency and, at the same time, some degree have three classes of high, middle, and low income fami-
of difference. More specifically, for objects with evaluation lies. In conflict management, one may consider three levels
values close to the four corners of the square, the two of low, medium/moderate, and high conflict (Yao 2019a).
methods are consistent. They are different when evaluation It is also possible to interpret the interval [0, 1] as a
bipolar scale with 0 and 1 representing, respectively, the
maximum negativity and the maximum positivity. Under
(+1, −1) this interpretation, we have L½0;l ðeÞ, M ðl;hÞ ðeÞ, and H ½h;1 ðeÞ,
which may also be interpreted as NEG½0;l ðeÞ, BNDðl;hÞ ðeÞ,
en + ep = l h = en + ep and POS½h;1 ðeÞ.

NEG BND POS

(0, −1) (+1, 0) 4 Three-way approximations
of nonstandard sets
ne

ep
ga

ity
tiv

A main advantage of three-way decision is the simplicity

iv
ity

sit
po

offered by thinking in threes. Many nonstandard sets typ-

en

(0, 0) ically use many-valued and infinite-valued membership

functions, which may lead to difficulties in understanding
Fig. 4 Trisection based an aggregated evaluation

123
 Granular Computing

and processing. By interpreting membership functions of description language. That is, knowledge as represented by
nonstandard sets as evaluations, we can obtain their three- the description language is granulated in the sense that we
way approximations based on evaluation-based models of have to consider an equivalence class as one indecom-
three-way decision introduced in the last section. posable unit instead of many individual objects. Equiva-
lence classes are the minimal non-empty definable sets
4.1 Three-way decision with rough sets (Yao 2015). We can only study these equivalence granules
by using the description language. In fact, it can be easily
Pawlak (1982, 1991) introduced the notion of a rough set as verified that the three regions can be equivalently defined
an approximation of an indefinable set by using a pair of as unions of some equivalence classes as follows:
definable sets called the lower and upper approximations.  
POSðXÞ ¼ [ ½xE 2 U=E j ½xE  X ;
A set is definable if it can be described by a formula in a  
BNDðXÞ ¼ [ ½xE 2 U=E j ½xE \ X 6¼ ; ^ :ð½xE  XÞ ;
description language and is indefinable otherwise (Yao  
2015). The lower and upper approximations can be NEGðXÞ ¼ [ ½xE 2 U=E j ½xE \ X ¼ ; :
equivalently expressed as three pairwise disjoint definable ð6Þ
sets, namely the positive region defined by the lower
An equivalence class is completely included in exactly one
approximation, the boundary region defined by the differ-
of the three regions.
ence of the upper and the lower approximations, and the
In light of evaluation-based model of three-way deci-
negative region defined by the complement of the upper
sion, we define an evaluation eðjXÞ : U ! ½0; 1 by using
approximation. The three regions form a disjoint trisection
the following conditional probability of the probabilistic
for three-way decision.
rough set model (Yao 2010):
Let E denote an equivalence relation on a finite uni-
versal set of objects U. The pair apr ¼ ðU; EÞ is called an jX \ ½xE j
eðxjXÞ ¼ PrðXj½xE Þ ¼ ; ð7Þ
approximation space. The equivalence class containing an j½xE j
object x 2 U is given by:
where j  j denotes the cardinality of a set. It follows that
½xE ¼ fy 2 U j xEyg: ð4Þ 0  eðxjXÞ  1. Moreover, objects in the same equivalence
class have the same value, that is, if y 2 ½xE , then
The family of all equivalence classes forms a partition
eðyjXÞ ¼ eðxjXÞ. For a pair of thresholds (l, h) with
U=E ¼ f½xE j x 2 Ug of U. We may use a set of attributes
0  l\h  1, by inserting the valuation eðxjXÞ ¼
to construct a description language for describing the set of
PrðXj½xE Þ into Definition 4, we immediately have a defi-
objects. The equivalence relation E is defined by a set of
nition of probabilistic three-way approximations.
attributes so that two objects are equivalent if they have the
same values on the set of attributes. In other words, two Definition 6 In an approximation apr ¼ ðU; EÞ, with
objects are equivalent if they satisfy the same set of for- respect to the conditional probability PrðXj½xE Þ ¼ jX \
mulas in the description language. The empty set, equiva- ½xE j=j½xE j and a pair of thresholds (l, h) with 0  l\h  1,
lence classes of E, and unions of families of equivalence the three probabilistic positive, boundary, and negative
classes are definable sets. Equivalence classes are ele- regions are defined as follows:
mentary definable sets used to construct rough set  
approximations. POS½h;1 ðXÞ ¼ x 2 U j PrðX j ½xE Þ h
 
¼ [ ½xE 2 U=E j PrðX j ½xE Þ h ;
Definition 5 In an approximation space apr ¼ ðU; EÞ, a  
three-way approximation of a subset of object X  U is BNDðl;hÞ ðXÞ ¼ x 2 U j l\PrðXj½xE Þ\h
  ð8Þ
given as follows: ¼ [ ½xE 2 U=E j l\PrðXj½xE Þ\h ;
   
POSðXÞ ¼ x 2 U j ½xE  X ; NEG½0;l ðXÞ ¼ x 2 U j PrðXj½xE Þ  l
   
BNDðXÞ ¼ x 2 U j ½xE \ X 6¼ ; ^ :ð½xE  XÞ ; ð5Þ ¼ [ ½xE 2 U=E j PrðXj½xE Þ  l :
 
NEGðXÞ ¼ x 2 U j ½xE \ X ¼ ; :

The triplet hPOSðXÞ; BNDðXÞ; NEGðXÞi is called a three- Again, an equivalence class is completely included in
way rough set approximation of X. exactly one of the three regions. By setting l ¼ 0 and
An important notion of rough set theory is the knowl- h ¼ 1, we obtain the standard rough set approximation in
edge granularity caused by a description language. All Definition 5, that is, POSðXÞ ¼ POS½1;1 ðXÞ,
objects in an equivalence class have the same description, BNDðXÞ ¼ BNDð0;1Þ ðXÞ, and NEGðXÞ ¼ NEG½0;0 ðXÞ. By
and it is impossible to differentiate them according to the our formulation, rough sets and probabilistic rough sets can

123
Granular Computing

be interpreted as set-theoretic models of three-way deci- intervals of f0; 1g, where ½0; 0 ¼ f0g; ½0; 1 ¼ f0; 1g; ½1; 1
sion. The universal set U is trisected into three regions to ¼ f1g, and ½0; 0  ½0; 1  ½1; 1. We can immediate for-
approximate a set. mulate an evaluation by using the following membership
function of an interval set:
4.2 Three-way decision with interval sets 8
>
< ½0; 0; x 2 ðXÞc ;
The notion of an interval set was introduced by Yao eðxj½X; XÞ ¼ ½0; 1; x 2 X  X; ð11Þ
>
:
(1993, 2017) as a means to describe partially known con- ½1; 1; x 2 X:
cepts. On the one hand, it is assumed that an object is either
This also provides an interpretation of an interval set as an
a member or not a member of a set. On the other hand, due
interval-valued set, which is a generalization of single-
to a lack of information and knowledge, we can only
valued interpretation of a set. If we know that an object is a
express the state of member and non-members for some
member of a set, its membership value is [1, 1], which is
objects, instead of all objects. That is, we have a partially
equivalent to 1. If we know that an object is not a member
known set defined by a lower bound and an upper bound.
of the set, its membership value is [0, 0], which is equiv-
The lower bound consists of objects that are known to be
alent to 0. When we do not know the actual state of an
members of a set, the complement of the upper bound
object, although we cannot precisely give its membership
consists of objects that are known to be non-members of
value, we do know that the actual membership value is
the set, and the actual states of the rest of objects are
either 0 or 1. In this case, ½0; 1 ¼ f0; 1g is used.
unknown. Although a pair of lower and upper bounds of an
For notional simplicity, we rewrite the totally order set
interval set is similar in form to a pair of lower and upper
ðf½0; 0; ½0; 1; ½1; 1g; Þ as ðf0; u; 1g; Þ. By using the
approximations of a rough set, they have very different
evaluation (11) and setting l ¼ 0 and h ¼ 1 in Definition 3,
semantics interpretations. While a rough set is an approx-
we have another definition of the three regions of an
imation of an indefinable set whose membership function is
interval set.
given, an interval set is an approximation of a partially
known set whose membership function is not fully given. Definition 8 Based on the evaluation (11), the negative,
boundary, and positive regions of an interval set are
Definition 7 Let U be a finite non-empty universal set of
equivalently defined by:
objects and 2U be the power set of U. For a pair of subsets  
of objects X and X with X  X, a closed interval set of U is NEGð½X; XÞ ¼ x 2 U j eðxj½X; XÞ40 ;
 
a subset of 2U defined as follows: BNDð½X; XÞ ¼ x 2 U j 0  eðxj½X; XÞ  1 ; ð12Þ
   
½X; X ¼ Y 2 2U j X  Y  X : ð9Þ POSð½X; XÞ ¼ x 2 U j eðxj½X; XÞ<1 :

The two subsets X and X are called the lower and upper
bounds of the interval set, respectively. The re-expression of an interval set in terms of three-
The definition of an interval set as a family of sets gives way decision is straightforward. The purpose is to show
rise to a possible-world semantics. Each set in the interval that interval sets can be formulated in the framework of
set may be the actual set of the partially known set in a three-way decision. It should also be pointed out that the
three-valued definition of an interval set hides, unfortu-
possible world. Alternatively, an interval set ½X; X can be
nately, the possible-world semantics as given by the defi-
equivalently defined by three regions:
nition in terms of a family of sets.
NEGð½X; XÞ ¼ ðXÞc ;
BNDð½X; XÞ ¼ X  X; ð10Þ 4.3 Three-way decision with fuzzy sets
POSð½X; XÞ ¼ X: and shadowed sets

This definition offers a three-way interpretation of an Zadeh (1965) proposed the notion of a fuzzy set for
interval set. Objects in NEGð½X; XÞ and POSð½X; XÞ are describing a concept with a gradually changing unsharp
known to be non-members and members of the partially boundary.
known set. The states of objects in BNDð½X; XÞ are
Definition 9 A fuzzy set on a finite non-empty universal
unknown. For a set, its characteristic function is from U to
set of objects U is defined by a membership function
ðf0; 1g; Þ with 0  1, where 0 and 1 indicate, respec-
lA : U ! ½0; 1. For x 2 U, lA ðxÞ 2 ½0; 1 is the mem-
tively, non-member and member of a set. To have a three-
bership value/grade of x.
valued interpretation of an interval set, we consider the
totally ordered set, ðf½0; 0; ½0; 1; ½1; 1g; Þ, formed by all

123
 Granular Computing

As suggested by Dubois and Prade (2012), gradualness a shadowed set as a three-way approximation of lA as
is only one of several possible semantics interpretations of follows:
a fuzzy set (Dubois 2010). Three-way decision with fuzzy 8
>
< 0; lA ðxÞ  l;
sets is based on the gradualness interpretation. The mem-
SA ðxÞ ¼ ½0; 1; l\lA ðxÞ\h; ð14Þ
bership value/grade indicates the degree to which an object >
:
belongs to a fuzzy set, or the degree to which the object is 1; lA ðxÞ h:
an instance of the concept modeled by a fuzzy set. Mem-
bership grades of objects provide quantitative information
about objects. In many situations, it may be more mean- The construction process of a shadowed set may be
ingful to consider qualitative information provided by interpreted based on three actions (Deng and Yao 2014;
membership grades. For example, an object with the Pedrycz 1998; Pedrycz and Vukovich 2002; Pedrycz
highest membership grade 1 is a full member and an object 2009): (a) if the fuzzy membership grade of an object x is
with the lowest membership grade 0 is a full non-member. close to 1, i.e., lA ðxÞ h, we lift lA ðxÞ to 1, (b) if the fuzzy
The use of the two extreme values may be too restrictive to membership grade is close to 0, i.e., lA ðxÞ  l, we reduce
be practically useful. It might be reasonable to treat lA ðxÞ to 0, and if the fuzzy membership grade is in the
membership grades close to 1 as if they are 1 and to treat middle, i.e., l\lA ðxÞ  h, we extend lA ðxÞ to the unit
membership grades close to 0 as if they are 0. In fact, interval [0, 1]. For representing a shadowed set, Pedrycz
Zadeh (1965) made this suggestion in his seminal paper on (1998) used the totally ordered set ðf0; ½0; 1; 1g; Þ with
fuzzy sets, which is quoted here: 0  ½0; 1  1. While the membership grades of 0 and 1
represent the least uncertainty, membership grade [0, 1]
‘‘...one can introduce two levels a and b (0\a\1,
represents the most uncertainty. Semantically, it may be
0\b\1, a [ b) and agree to say that (1) ‘x belongs
more accurate to use a totally ordered set
to A’ if fA ðxÞ a; (2) ‘x does not belong to A’ if
ðf½0; l; ðl; hÞ; ½h; 1g; Þ with ½0; l  ðl; hÞ  ½h; 1 to
fA ðxÞ  b; and (3) ‘x has an indeterminate status rel-
define a shadowed set. The three intervals, [0, l], (l, h), and
ative to A’ if b\fA ðxÞ\a. This leads to a three-val-
[h, 1], indicate the ranges in which the membership grade
ued logic (Kleene, 1952) with three truth values: T
of an element lies.
(fA ðxÞ a), F (fA ðxÞ  b), and U (b\fA ðxÞ\a).’’
In general, as suggested by Yang and Yao (2020), we
Zadeh’s idea motivates the introduction of three-way may use a totally ordered set ðfw; g; bg; Þ with w  g 
approximations of a fuzzy set (Deng and Yao 2014; Yao b for defining a shadowed set. The three values represent,
et al. 2017). respectively, the white, gray, and black members of a
In light of three-way decision with one evaluation, by shadowed. By using SA as an evaluation function and
treating lA as an evaluation in Definition 4, i.e., setting l ¼ w and h ¼ b in Definition 3, we have a defi-
eðxjlA Þ ¼ lA ðxÞ, we immediately have three-way nition of the three regions of a shadowed set.
approximations of a fuzzy set. Definition 12 The negative (i.e., white), boundary (i.e.,
Definition 10 Suppose lA : U ! ½0; 1 is a fuzzy set on a shadowed or gray), and positive (i.e., black) regions of a
finite non-empty set of objects U. Given a pair of thresh- shadowed set are defined by:
olds (l, h) with 0  l\h  1, we define the following three- NEGðSA Þ ¼ fx 2 U j SA ðxÞ4wg;
way approximations of lA :
BNDðSA Þ ¼ fx 2 U j w  SA ðxÞ  bg; ð15Þ
NEG½0;l ðlA Þ ¼ fx 2 U j lA ðxÞ  lg; POSðSA Þ ¼ fx 2 U j SA ðxÞ<bg:
ðl;hÞ
BND ðlA Þ ¼ fx 2 U j l\lA ðxÞ\hg; ð13Þ
½h;1
POS ðlA Þ ¼ fx 2 U j lA ðxÞ hg:
A shadowed set shares the same three-valued form as an
interval set. However, they have very different semantics
interpretations. An interval set is a three-way representa-
Pedrycz (1998, 2009) introduced the notion of a shad-
tion of a partially known set, a shadowed set is a three-way
owed set as a three-way or a three-valued approximation of
approximation of a fuzzy set with a gradually changing
a fuzzy set (Deng and Yao 2014; Pedrycz and Vukovich
boundary.
2002; Yao et al. 2017).
Definition 11 Suppose lA : U ! ½0; 1 is a fuzzy set
defined on a finite non-empty set U. Given a pair of
thresholds (l, h) with 0  l\h  1, Pedrycz (1998) defined

123
Granular Computing

4.4 Three-way decision with rough fuzzy sets membership grades in the middle. Since not all values in
the interval ½lA ðxÞ; lA ðxÞ are used by lA for objects in
Dubois and Prade (1990) introduced the notion of a rough ½xA , it may not be reasonable to use the middle point of the
fuzzy set to study approximation of a fuzzy set in an interval. Along the same line of constructing probabilistic
approximation space. As explained earlier, in an approxi- rough sets, we may consider the following aggregated
mation space apr ¼ ðU; EÞ we have to treat an equivalence membership function:
class ½xE as an indecomposable unit rather than a collec- P
þ y2½xE lA ðyÞ
tion of individuals. Consider a fuzzy set lA . Although eðxjlA Þ ¼ lA ðxÞ ¼ : ð17Þ
objects in ½xE may have different membership grades, j½xE j
P
when ½xE is considered as an indecomposable unit, we are The value y2½xE lA ðyÞ may be interpreted as the cardi-
required to use the same membership grade for all objects nality or sigma-count of the intersection of set ½xE and
in ½xE . On the other hand, we cannot use a single grade fuzzy set lA . In fact, if the fuzzy set lA is a set with
because membership grades of objects in ½xE may be dif- membership grades from f0; 1g, Eq. (17) reduces to
ferent. To solve this problem, we use a pair of fuzzy sets as Eq. (7) of probabilistic rough sets, which suggests that the
lower and upper approximations of the fuzzy set. aggregated membership function is reasonable. It follows
Definition 13 Suppose that apr ¼ ðU; RÞ is an approxi- that 0  eðxjlA Þ  1. If y 2 ½xE , then lþ þ
A ðyÞ ¼ lA ðxÞ.
þ
mation space defined by an equivalence relation E on a With the membership function lA , we transform a fuzzy
finite non-empty set of objects U. For a fuzzy set lA , a pair set lA into a fuzzy set lþ A to reflect knowledge granularity
of its lower and upper approximations is given by: caused by an approximation space. By inserting fuzzy set
  lþ
lA ðxÞ ¼ min lA ðyÞ j y 2 ½xE ; A into Definition 10, we obtain three regions of a fuzzy
  ð16Þ set in an approximation space.
lA ðxÞ ¼ max lA ðyÞ j y 2 ½xE :
Definition 14 Suppose that apr ¼ ðU; EÞ is an approxi-
The pair of sets is called a rough fuzzy set induced by the mation space defined by an equivalence relation E on a
fuzzy set lA . finite non-empty set of objects U. Let a fuzzy set lþ A,
By definition, all objects in ½xE have the same lower and constructed by Eq. (17), denote the approximation of a
upper membership grades as x, that is, if y 2 ½xE , then fuzzy set lA : U ! ½0; 1 in apr ¼ ðU; EÞ. Given a pair of
lA ðyÞ ¼ lA ðxÞ and lA ðyÞ ¼ lA ðxÞ. In other words, in the thresholds (l, h) with 0  l\h  1, we define the following
lower and upper membership functions, we are able to three-way approximations of lA :
 
consider an equivalence class as an indecomposable unit. NEG½0;l ðlA Þ ¼ NEG½0;l ðlþ þ
A Þ ¼ x 2 U j lA ðxÞ  l ;
As a matter of fact, Dubois and Prade (1990) originally  
BNDðl;hÞ ðlA Þ ¼ BNDðl;hÞ ðlþ þ
A Þ ¼ x 2 U j l\lA ðxÞ\h ;
defined a fuzzy rough set as a pair of fuzzy sets on the  
quotient space U/E with lA ð½xE Þ ¼ lA ðxÞ and POS½h;1 ðlA Þ ¼ POS½h;1 ðlþ þ
A Þ ¼ x 2 U j lA ðxÞ h :
lA ð½xE Þ ¼ lA ðxÞ. We use Definition 13 with fuzzy sets on ð18Þ
U in order to be consistent to other types of generalizations
of sets.
For any object x 2 U, we have lA ðxÞ  lA ðxÞ  lA ðxÞ. Similarly, we can also define a shadowed set as an
Mathematically, one may use the interval ½lA ðxÞ; lA ðxÞ as approximation of the fuzzy set lA in an approximation
the membership grade of x to represent a rough fuzzy set. space apr ¼ ðU; EÞ.
That is, a rough fuzzy set has the same form of an interval- In constructing a three-way approximation of a fuzzy set
valued fuzzy set to be discussed in the next subsec- in an approximation space, we use the aggregated fuzzy
tion. However, their semantics interpretations are very membership function lþ A instead of the pair of lower and
different. An important observation is that for an arbitrary upper fuzzy sets. Our formulation is in line with the for-
number v 2 ½lA ðxÞ; lA ðxÞ there may not exist an object mulation of probabilistic rough sets. With the pair of lower
and upper fuzzy sets, it may be attempting to applying the
y 2 ½xE such that lA ðyÞ ¼ v. In fact, at most j½xE j number
model with two evaluations in which lA is treated as the
of values in the interval ½lA ðxÞ; lA ðxÞ are used by lA for
positive evaluation and ð1  lA Þ as the negative evalu-
objects in ½xE .
ation. As demonstrated earlier, the pair of lower and upper
When constructing a rough fuzzy set, we only consider
fuzzy sets does not take sufficient consideration of mem-
the two extreme cases, namely objects in ½xE with the
bership grades of objects in ½xE . We therefore do not take
smallest and the largest membership grades. In the spirit of
this approach. As will be shown in the next subsection, a
three-way decision, we also need to consider objects with

123
 Granular Computing

formulation with two evaluations may be appropriate for That is, each fuzzy set in the family corresponds to one
building three-way decision models with interval fuzzy possible world in which the fuzzy set is the actual and true
sets. fuzzy set. Third, our possible-world semantics is only one
of many possible interpretations. It might be worthwhile to
4.5 Three-way decision with interval fuzzy sets investigate other interpretations.
In light of evaluation-based three-way decision, Yang
Interval fuzzy sets are a generalization of fuzzy sets. In this and Yao (2019) introduced a method for constructing a
case, we have both ontic and epistemic uncertainties. On three-way approximation of an intuitionistic fuzzy set. We
the one hand, a concept has a non-sharp and gradually reformulate their results by using the notations of interval
changing boundary, which is modeled by a fuzzy set. On fuzzy sets. We interpret 0  lA  1 as a positive evalua-
the other hand, we also do not have fully knowledge about tion, i.e., ep ðxj½lA ; lA Þ ¼ lA ðxÞ, and 1   ð1  lA Þ
the actual and true membership function. Instead, we only  0 as a negative evaluation, i.e., en ðxj½lA ; lA Þ ¼
know a pair of a lower bound and an upper bound within
ð1  lA ðxÞÞ. By setting ðLþ ; Þ and ðL ; Þ, respec-
which lies the true membership function, similar to the
interpretation of an interval set. tively, as ð½0; 1; \Þ and ð½1; 0; \Þ in Definition 2, we
define a three-way approximations of an interval fuzzy set.
Definition 15 Suppose that U is a finite non-empty set of
objects. For two fuzzy sets lA and lB on U, we say that lA Definition 16 Suppose that ½lA ; lA  is an interval fuzzy
is a fuzzy subset of lB , written lA  lB if for every x 2 U, set on a finite non-empty set of objects U. Given a pair of
lA ðxÞ  lB ðxÞ. An interval fuzzy set ½lA ; lA  on U is thresholds htp ; tn i with 1  tn  0 and 0  tp  1, the three
defined by a pair of fuzzy sets lA and lA with lA  lA regions of the interval fuzzy set are defined by:
h i
and is interpreted as the following family of fuzzy sets: NEGhtp ;tn i lA ; lA
n o n o
½lA ; lA  ¼ lA j lA  lA  lA : ð19Þ ¼ x 2 U j ð1  lA ðxÞÞ  tn ^ lA ðxÞ\tp
n o
¼ x 2 U j 1  lA ðxÞ  tn ^ lA ðxÞ\tp ;
h i
Several comments on the notion of an interval fuzzy set POShtp ;tn i lA ; lA
are given as follows. First, by interpreting an interval fuzzy n o
set as a family of fuzzy sets, we are able to represent both ¼ x 2 U j ð1  lA ðxÞÞ [ tn ^ lA ðxÞ tp
ontic and epistemic uncertainties. It is assumed that exactly n o
one of the fuzzy sets in the family is the actual and true ¼ x 2 U j 1  lA ðxÞ\  tn ^ lA ðxÞ tp ;
fuzzy set, if our knowledge or information becomes com- h i
BNDhtp ;tn i lA ; lA
plete. Second, an interval fuzzy set can be shown to be
 h ic
mathematically equivalent to a few other notions. One may ¼ NEGhtp ;tn i ð½lA ; lA Þ [ POShtp ;tn i lA ; lA
treat the interval ½lA ðxÞ; lA ðxÞ as an interval-valued grade n o
of x. This leads to the notion of interval-valued type two ¼ x 2 U j ð1  lA ðxÞÞ  tn ^ lA ðxÞ tp [
fuzzy sets as another representation of interval fuzzy sets n o
(Dubois 2010; Dubois and Prade 2012). By the facts that y 2 U j ð1  lA ðyÞÞ [ tn ^ lA ðyÞ\tp
n o
0  lA ðxÞ  1, 0  lA ðxÞ  1, and lA ðxÞ  lA ðxÞ, we can
¼ x 2 U j 1  lA ðxÞ  tn ^ lA ðxÞ tp [
conclude that 0  1  lA ðxÞ  1 and n o
0  lA ðxÞ þ ð1  lA ðxÞÞ  1. If we interpret the value y 2 U j 1  lA ðyÞ\  tn ^ lA ðyÞ\tp :
lA ðxÞ as representing membership, and the value 1 
lA ðxÞ as representing non-membership, the pair
hlA ðxÞ; 1  lA ðxÞi immediately gives rise to the notion of As shown in Fig. 3, the boundary region consists of two
an intuitionistic fuzzy set (Atanassov 1986, 1999) or the parts. The first part fx 2 U j 1  lA ðxÞ  tn ^
notion of a vague set (Bustince and Burillo 1996; Gau and lA ðxÞ tp g is formed by objects with both high member-
Buehrer 1993). Although these notions are mathematically ship grades, i.e., lA ðxÞ tp , and high non-membership
equivalent, they may have different semantics interpreta- grades, i.e., 1  lA ðxÞ  tn . The second part fy 2 U j
tions. They may also offer different hints for different 1  lA ðyÞ\  tn ^ lA ðyÞ\tp g consists of objects with
applications. In this paper, we adopt the interpretation of an
both low membership grades and low non-membership
interval fuzzy set as a family of fuzzy sets. Like the case of
grades.
interval sets, we in fact have a possible-world semantics.

123
Granular Computing

Yang et al. (2019b) recently proposed another method characterized, respectively, lA ðxÞ þ lA ðxÞ  1 ¼ l and
for constructing three-way approximation of an interval lA ðxÞ þ lA ðxÞ  1 ¼ h.
fuzzy set. Their formulation is based on the membership
and non-membership functions of an intuitionistic fuzzy 4.6 Three-way decision with soft sets
set. We review their results by using the notation of
interval fuzzy sets. According to Eq. (2), we can combine Molodtsov (1999) introduced the notion of soft sets for
the pair of evaluations lA and ð1  lA Þ into one representing concepts under uncertainty. Formally, a soft
evaluation: set is defined by a parameterized family of sets. For sim-
eðxj½lA ; lA Þ ¼ lA ðxÞ þ ðð1  lA ðxÞÞÞ plicity, we only consider soft sets based on a finite non-
ð20Þ empty set of parameters. In this subsection, we briefly
¼ lA ðxÞ þ lA ðxÞ  1:
review a method suggested by Yang and Yao (2020) for
We have 1  eðxj½lA ; lA Þ  þ 1. That is, eðxj½lA ; lA Þ constructing a three-way approximation of a soft set.
is an evaluation using the scale ð½1; þ1; \Þ. Definition 18 Let U be a finite non-empty universal set of
Furthermore, objects and P be a finite non-empty set of parameters. The
power set of U, denoted by 2U , consists of all subsets of U.
eðxj½lA ; lA Þ ¼ 1 () lA ðxÞ ¼ lA ðxÞ ¼ 0;
ð21Þ A soft set over U is a pair (S, E), where E  P is a subset of
eðxj½lA ; lA Þ ¼ þ1 () lA ðxÞ ¼ lA ðxÞ ¼ 1: parameters and
By inserting this evaluation into Definition 4, we imme- S : E ! 2U ; ð23Þ
diately derive another three-way approximation of an
is a mapping from E to the power set of U. For any e 2 E,
interval fuzzy set.
SðeÞ  U is called an e-defined set of objects of (S, E).
Definition 17 Suppose that ½lA ; lA  is an interval fuzzy
Yang and Yao (2020) presented a conceptualization of
set on a finite non-empty set of objects U. Given a pair of soft sets that is slightly different from the original formu-
thresholds (l, h) with 1  l\0\h  þ 1, the three lation by Molodtsov. They suggested two semantics of soft
regions of the interval fuzzy set are defined as follows: sets, namely a multi-context semantics and a possible-
n o
NEG½1;l ð½lA ; lA Þ ¼ x 2 U j lA ðxÞ þ lA ðxÞ  1  l ; world semantics. In the multi-context semantics, the set of
n o parameters is interpreted as contexts in which a concept is
BNDðl;hÞ ð½lA ; lA Þ ¼ x 2 U j l\lA ðxÞ þ lA ðxÞ  1\h ; defined. In a specific context, the concept is precise and its
n o extension is a subset of objects from U. That is, there is no
POS½h;þ1 ð½lA ; lA Þ ¼ x 2 U j lA ðxÞ þ lA ðxÞ  1 h :
uncertainty in each context. The ontic uncertainty comes
ð22Þ from the fact that the same concept is given by different
subsets of objects in different contexts. That is, if we pool
all contexts together, the concept becomes uncertain. The
Figure 5 illustrates a trisection of the evaluation space possible-world semantics is similar to the possible-world
used for constructing a three-way approximation of an semantics of an interval set. In each possible world, a
interval fuzzy set. In this case, due to the condition concept is precisely given by a subset of objects. However,
0  lA ðxÞ  lA ðxÞ  1, the evaluation space is a triangle our knowledge or available information is insufficient for
instead of a square as in Fig. 4. The three regions are us to determine the possible world that we are in. In this
obtained from an aggregated evaluation case, soft sets provide a tool to represent epistemic
lA ðxÞ þ lA ðxÞ  1. The dotted line is characterized by uncertainty.
Consider a parameter e 2 E of a soft set (S, E), where
lA ðxÞ þ lA ðxÞ  1 ¼ 0. Two dashed lines are
E 6¼ ;. For an object x 2 U, if x 2 SðeÞ, we say that x is an
instance of a concept represented by (S, E) under the
l 0 h parameter e. Intuitively speaking, the number of parame-
(0, −1) (+1, 0) ters under which x is an instance of the concept provides a
NEG BND POS
reasonable evaluation. A probabilistic version is given by:
−
(1

jfe 2 E j x 2 SðeÞgj
A

eðxjðS; EÞÞ ¼ ð24Þ

−

:
µ

jEj
µA
)

(0, 0) We have 0  eðxjðS; EÞÞ  1, eðxjðS; EÞÞ ¼ 0 if and only if

x is not an instance under any of the parameters in E, and
Fig. 5 Trisection used for approximating interval fuzzy sets

123
 Granular Computing

eðxjðS; EÞÞ ¼ 1 if and only if x is an instance under all Our formulations are based on evaluation-based models
parameters in E. of three-way decision, in which a pair of evaluations or a
The evaluation e(x|(S, E)) is defined with respect to the single evaluation is used. We have reported a detailed
scale ð½0; 1; \Þ. By inserting the evaluation e(x|(S, E)) analysis of an evaluation space defined by a pair of eval-
into Definition 4, we immediately have a three-way uations. The construction of a three-way approximation of
approximation of a soft set. a nonstandard set is a two-step process: (1) define evalu-
ation(s) based on information given by a nonstandard set,
Definition 19 Suppose that (S, E) is a soft set on a finite
and (2) apply an evaluation-based model to trisect the
non-empty set of objects U and a finite non-empty set of
universal set of objects. For different types of nonstandard
parameters P. Given a pair of thresholds (l, h) with
sets, we have presented different ways to define evalua-
0  l\h  1, the three regions of the soft set are defined as
tion(s). It perhaps should be pointed out that our formula-
follows:
tions depend on a specific interpretation of a nonstandard
NEG½0;l ððS; EÞÞ ¼ fx 2 U j eðxjðS; EÞÞ  lg; set. In other words, there are other different ways to define
BNDðl;hÞ ððS; EÞÞ ¼ fx 2 U j l\eðxjðS; EÞÞ\hg; ð25Þ evaluation(s) to reflect other semantics interpretations of a
nonstandard set. This is an area that demands further in-
POS½h;1 ððS; EÞÞ ¼ fx 2 U j eðxjðS; EÞÞ hg; depth investigation.
where eðxjðS; EÞÞ ¼ jfe 2 E j x 2 SðeÞgj=jEj. When constructing a trisection of the set of objects, we
rely on the use of thresholds on evaluation(s). In this paper,
If we interpret e(x|(S, E)) as the membership function of we have not discussed about methods to compute or set the
a fuzzy set, then Definition 19 is in fact the same as Def- required thresholds. There have been extensive research on
inition 10. In this way, one may consider the construction this topic (Afridi et al. 2018; Deng and Yao 2014; Pedrycz
of a three-way approximation of a soft set as a two-step 1998; Qiao and Hu 2020; Yang and Yao 2019, 2020; Yao
process. The first step transforms a soft set into a fuzzy set, 2010; Yao et al. 2017). Determining the required thresh-
and the second step constructs a three-way approximation olds is another area of worthwhile future research.
of the resulting fuzzy set (Yang and Yao 2020). In com- By building set-theoretic models of three-way decision,
parison, the probabilistic interpretation of the evaluation we have in fact demonstrated the power of the general
e(x|(S, E)) seems to be more appropriate than the fuzzy set TAO (trisecting–acting–outcome) framework of three-way
interpretation. decision. As a third area of future research, we may look at
new models of three-way decision. Many researchers have
already reported valuable results (Horiuchi et al. 2020; Hu
5 Conclusion remarks 2014; Hu et al. 2017; Li et al. 2017a, b, 2018; Li 2019;
Liang et al. 2020; Liu and Liang 2016; Liu et al. 2016;
In the context of concept analysis, we have examined six Yang et al. 2019a; Yu 2018; Zhang et al. 2020; Zhang and
models of three-way decision with nonstandard sets. We Yao 2020). These studies serve as a solid basis for new
assume that a set, either standard or nonstandard, repre- investigations.
sents the extension of a concept. Nonstandard sets can
represent two types of uncertainty. Ontic uncertainty is Acknowledgements This work was supported in part by a Discovery
captured by a concept with a gradually changing boundary, Grant from NSERC, Canada. The author thanks Professors Witold
and epistemic uncertainty is captured by either the inde- Pedrycz and Shyi-Ming Chen for their encouragements during the
preparation of the paper. The author is grateful to reviewers for their
finability of the intension of a concept or our inability to constructive and critical comments.
specify precisely the extension of a concept. With respect
to these specific understandings of various nonstandard Compliance with ethical standards
sets, we discuss their three-way approximations. A three-
way approximation enables us to accept or reject, beyond Conflict of interest There is no conflict of interest of the research
reasonable doubt under uncertainty, some objects to be reported in this paper.
instances of a concept. At the same time, we make non-
commitment decision for the rest of objects. Set-theoretic
models of three-way decision provide effective and prac- References
tical methods and tools for decision-making under uncer-
Afridi MK, Azam N, Yao JT, Alanazi E (2018) A three-way
tainty. Although our discussions in this paper are restricted clustering approach for handling missing data using GTRS. Int J
to concept formation and representation, the ideas of three- Approx Reason 98:11–24
way decision can be easily applied to other problems
involving uncertainty.

123
Granular Computing

Agbodah K (2019) The determination of three-way decisions with Li ZW, Huang D (2019) A three-way decision method in a fuzzy
decision-theoretic rough sets considering the loss function condition decision information system and its application in
evaluated by multiple experts. Granul Comput 4:285–297 credit card evaluation. Granul Comput. https://doi.org/10.1007/
Arnauld A, Nicole P (1996) Logic or the art of thinking, translated by s41066-019-00172-8
Jill Vance Buroker. Cambridge University Press, Cambridge Liang DC, Wang MW, Xu ZS, Liu D (2020) Risk appetite dual
Atanassov KT (1986) Intuitionistic fuzzy sets. Fuzzy Sets Syst hesitant fuzzy three-way decisions with TODIM. Inf Sci
20:87–96 507:585–605
Atanassov KT (1999) Intuitionistic fuzzy sets. Springer, Heidelberg Liu D, Liang DC (2016) Generalized three-way decisions and special
Béziau JV (2012) The power of the hexagon. Log Univ 6:1–43 three-way decisions. J Front Comput Sci Technol 11:502–510
Bustince H, Burillo P (1996) Vague sets are intuitionistic fuzzy sets. Liu D, Liang DC, Wang CC (2016) A novel three-way decision
Fuzzy Sets Syst 79:403–405 model based on incomplete information system. Knowl Based
Cacioppo JT, Berntson GG (1994) Relationship between attitudes and Syst 91:32–45
evaluative space: a critical review, with emphasis on the Ma JM, Zhang HY, Qian YH (2019) Three-way decisions with
separability of positive and negative substrates. Psychol Bull reflexive probabilistic rough fuzzy sets. Granul Comput
115:401–423 4:363–375
Cacioppo JT, Gardner WL, Berntson GG (1997) Beyond bipolar Mandal P, Ranadive AS (2019) Multi-granulation interval-valued
conceptualizations and measures: the case of attitudes and fuzzy probabilistic rough sets and their corresponding three-way
evaluative space. Pers Soc Psycholog Rev 1:3–25 decisions based on interval-valued fuzzy preference relations.
Cai MJ, Li QG, Lang GM (2017) Shadowed sets of dynamic fuzzy Granul Comput 4:89–108
sets. Granul Comput 2:85–94 Molodtsov D (1999) Soft set theory: first results. Comput Math Appl
Ciucci D (2011) Orthopairs: a simple and widely used way to model 37:19–31
uncertainty. Fundam Inform 108:287–304 Osgood CE, Suci GJ, Tannenbaum PH (1957) The measurement of
Couso I, Dubois D (2014) Statistical reasoning with set-valued meaning. University of Illinois Press, Chicago
information: ontic vs. epistemic views. Int J Approximate Pawlak Z (1982) Rough sets. Int J Comput Inform Sci 11:341–356
Reason 55:1502–1518 Pawlak Z (1991) Rough sets: theoretical aspects of reasoning about
Deng XF, Yao YY (2014) Decision-theoretic three-way approxima- data. Kluwer Academic Publishers, Boston
tions of fuzzy sets. Inf Sci 279:702–715 Pedrycz W (1998) Shadowed sets: representing and processing fuzzy
Dubois D (2010) Degrees of truth, ill-known sets and contradiction. sets. IEEE Trans Syst Man Cybern B Cybern 28:103–109
In: Bouchon-Meunier B, Magdalena L, Ojeda-Aciego M, Pedrycz W (2009) From fuzzy sets to shadowed sets: interpretation
Verdegay JL, Yager RR (eds) Foundations of reasoning under and computing. Int J Intell Syst 24:48–61
uncertainty. Springer, Berlin, pp 65–83 Pedrycz W, Vukovich G (2002) Granular computing with shadowed
Dubois D, Prade H (2008) An introduction to bipolar representations sets. Int J Intell Syst 17:173–197
of information and preference. Int J Intell Syst 23:866–877 Qiao JS, Hu BQ (2020) On decision evaluation functions in
Dubois D, Prade H (1990) Rough fuzzy sets and fuzzy rough sets. Int generalized three-way decision spaces. Inf Sci 507:733–754
J Gen Syst 17:191–209 Walker WE, Harremoës P, Rotmans J, van der Sluijs JP, van Asselt
Dubois D, Prade H (2012) Gradualness, uncertainty and bipolarity: MBA, Janssen P, Krayer von Krauss MP (2003) Defining
making sense of fuzzy sets. Fuzzy Sets Syst 192:3–24 uncertainty: a conceptual basis for uncertainty management in
Dundes A (1968) The number three in American culture. In: Dundes model-based decision support. Integr Assess 4:5–17
A (ed) Every man his way: readings in cultural anthropology. Yang JL, Yao YY (2020) Semantics of soft sets and three-way
Prentice-Hall, Englewood Cliffs, pp 401–424 decision with soft sets. Know Based Syst. https://doi.org/10.
Gau WL, Buehrer DJ (1993) Vague sets. IEEE Trans Syst Man 1016/j.knosys.2020.105538
Cybern 23:610–614 Yang JL, Yao YY (2019) From intuitionistic fuzzy sets to shadowed
Horiuchi K, Šešelja B, Tepavčević A (2020) Trice-valued fuzzy sets: sets: a three-way decision formulation. Manuscript
mathematical model for three-way decisions. Inf Sci Yang X, Li TR, Liu D, Fujita H (2019a) A temporal-spatial composite
507:574–584 sequential approach of three-way granular computing. Inf Sci
Hu BQ (2014) Three-way decisions space and three-way decisions. 486:171–189
Inf Sci 281:21–52 Yang XP, Li TJ, Tan AH (2019b) Three-way decisions in fuzzy
Hu BQ, Wong H, Yiu KFC (2017) On two novel types of three-way incomplete information systems. Int J Mach Learn Cybernet.
decisions in three-way decision spaces. Int J Approx Reason https://doi.org/10.1007/s13042-019-01025-1
82:285–306 Yao YY (1993) Interval-set algebra for qualitative knowledge
Lang GM, Luo JF, Yao YY (2020) Three-way conflict analysis: a representation. In: Proceedings of the fifth international confer-
unification of models based on rough sets and formal concept ence on computing and information, pp 370–374
analysis. Know Based Syst. https://doi.org/10.1016/j.knosys. Yao YY (2010) Three-way decisions with probabilistic rough sets. Inf
2020.105556 Sci 180:341–353
Li HX, Zhang LB, Zhou XZ, Huang B (2017a) Cost-sensitive Yao YY (2012) An outline of a theory of three-way decisions,
sequential three-way decision modeling using a deep neural RSCTC 2012. LNCS (LNAI) 7413:1–17
network. Int J Approx Reason 85:68–78 Yao YY (2015) The two sides of the theory of rough sets. Knowl
Li JH, Huang CC, Qi JJ, Qian YH, Liu WQ (2017b) Three-way Based Syst 80:67–77
cognitive concept learning via multi-granularity. Inf Sci Yao YY (2016) Three-way decisions and cognitive computing.
378:244–263 Cognit Comput 8:543–554
Li XN (2019) Three-way fuzzy matroids and granular computing. Int Yao YY (2017) Interval sets and three-way concept analysis in
J Approx Reason 114:44–50 incomplete contexts. Int J Mach Learn Cybernet 8:3–20
Li XN, Yi HJ, She YH, Sun BZ (2018) Generalized three-way Yao YY (2018) Three-way decision and granular computing. Int J
decision models based on subset evaluation. Int J Approx Reason Approx Reason 103:107–123
83:142–159 Yao YY (2019a) Three-way conflict analysis: reformulations and
extensions of the Pawlak model. Knowl Based Syst 180:26–37

123
 Granular Computing

Yao YY (2019b) Tri-level thinking: models of three-way decision. Int Zhang CY, Gao RY, Qin H, Feng XZ (2019b) Three-way clustering
J Mach Learn Cybernet. https://doi.org/10.1007/s13042-019- method for incomplete information system based on set-pair
01040-2 analysis. Granul Comput. https://doi.org/10.1007/s41066-019-
Yao YY (2020) Three-way granular computing, rough sets, and 00197-z
formal concept analysis. Int J Approx Reason 116:106–125 Zhang QH, Xia DY, Liu KY, Wang GY (2020) A general model of
Yao YY, Wang S, Deng XF (2017) Constructing shadowed sets and decision-theoretic three-way approximations of fuzzy sets based
three-way approximations of fuzzy set. Inf Sci 412–413:132–153 on a heuristic algorithm. Inf Sci 507:522–539
Yu H (2018) Three-way decisions and three-way clustering. In: Zhang Y, Yao JT (2020) Game theoretic approach to shadowed sets: a
IJCRS 2018, LNCS (LNAI), vol 11103, pp 13–28 three-way tradeoff perspective. Inf Sci 507:540–552
Zadeh LA (1965) Fuzzy sets. Inf Control 8:338–353
Zhang CY, Feng XZ, Gao RY (2019a) Three-way decision models Publisher’s Note Springer Nature remains neutral with regard to
and its optimization based on Dempster–Shafer evidence theory jurisdictional claims in published maps and institutional affiliations.
and rough sets. Granul Comput. https://doi.org/10.1007/s41066-
019-00201-6

123