The University of New South Wales
School of Electrical Engineering and Computer Science and Engineering
Mathematical Analysis of Some Rough Hybridized Models
Tanzeela Shaheen Doctor of Philosophy 2016
Supervisor: Prof. Dr. Muhammad Shabir Assessor: Dr. Abc ii Contents
1 Preliminaries 1 1.1 Rough sets ...... 1 1.1.1 Basic concepts ...... 1 1.1.2 Rough membership functions ...... 7 1.1.3 Generalizations of rough sets ...... 9 1.2 Fuzzy sets ...... 12 1.2.1 Operations on fuzzy sets ...... 13 1.3 Hesitant fuzzy sets ...... 14
2 Fuzzi ed Rough Sets 18 2.1 Introduction ...... 18 2.2 Fuzzi ed rough approximations for indiscernible objects ...... 19 2.3 Proposed Algorithm ...... ...... 22 2.3.1 Illustrative Example ...... 23 2.4 Properties of fuzzi ed rough approximations ...... 25 2.5 Fuzzi ed rough membership function ...... 28 2.6 Measures associated with fuzzi ed rough sets ...... 30 2.7 Generalization of fuzzi ed rough sets using logical connectives . . . 34 2.7.1 Proposed Algorithm ...... 37 2.8 Comparative analysis ...... 39 2.9 Conclusion ...... 43
3 Generalized Hesitant Fuzzy Rough Sets 44 3.1 Introduction ...... 44 3.2 Approximations of Generalized hesitant fuzzy rough sets (GHF RS) . 44 3.2.1 (GHF RS) approximation operators and its properties . . . . . 45 3.2.2 Singleton GHF R approximation operators ...... 51 3.3 Topological structure associated with GHF R sets and singleton GHF R sets...... 57 3.4 Three way decision making with GHF RS model ...... 63 3.4.1 Proposed algorithm ...... 64 3.4.2 An illustrative example ...... 66 3.4.3 Comparative analysis ...... 69 3.5 Conclusion ...... 73
i CONTENTS ii
4 Multi-granulation structure of fuzzi ed rough sets and GHFRS 74 4.1 Introduction ...... ...... 74 4.2 Multi-granulation rough sets ...... 74 4.3 Multi-granulation fuzzi ed rough ( MGFR) sets ...... 76 4.3.1 Illustrative example ...... ...... 76 4.3.2 Properties of MGF RS ...... 77 4.4 Multi-granulation singleton generalized hesitant fuzzy rough sets . . . 82 4.5 Conclusion ...... 87
5 Graphical equivalence relations and graphical partitions 89 5.1 Introduction ...... 89 5.2 Graphical Equivalence Relation and Graphical Partition ...... 89 5.2.1 Graphs and Hypergraphs ...... 89 5.2.2 Relations on Hypergraphs ...... 92 5.2.3 Graphical equivalence relations(GERs) ...... 93 5.2.4 Graphical Partitions(GPs) ...... 96 5.2.5 Correspondence between graphical equivalence relations and graph- ical partitions ...... 99 5.3 Graphical Partitions as Surjective Mappings ...... 100 5.4 Two-tier Graphical Partitions ...... 102 5.5 Conclusion ...... 107
6 Rough graphs and their algebraic structure 108 6.1 Introduction ...... 108 6.2 Rough graphs ...... 108 6.3 Algebraic structure of rough graphs ...... 113 6.4 Conclusion ...... 120 Chapter 1
Preliminaries
1.1 Rough sets
Classical set theory has been extended to many valuable set theories due to the ongoing interests in knowledge representation and analysis of incompleteness and uncertainty in data analysis. Among other theories, rough set theory has proved to be a worthwhile extension. As a mathematical tool, theory of rough sets was initially proposed by Pawlak [46, 47] to handle imprecision and incompleteness in information systems. The initial approach adopted by Pawlak includes partitioning the universe set into granules (classes) of elements which are indistinguishable or indiscernible subject to the available information. Using these classes, an arbitrary subset of a universe is approximated by two de nable subsets called lower and upper approximations. The concealed knowledge in the information system can thus be revealed by using the concepts of upper and lower approximations in the rough set theory.
1.1.1 Basic concepts
In this section, the basic notions, de nitions, and properties of rough sets are recalled. An example has been presented to demonstrate these concepts. Let U be a nite set of objects and R U U be a given equivalence relation. The sets U; R are called the universe and an indiscernilbility relation, respectively. The indiscernibility relation represents our lack of knowledge about elements of U. A pair (U; R) is called an approximation space, where U is the universe and R is an equivalence relation on U. Let X be a subset of U, i.e. X U. Our goal is to characterize the set X with respect to R. In order to do this, we need additional notation and basic concepts of rough set theory which are presented below.
By [x]R we denote the equivalence class of R determined by element x. The indis- cernilbility relation R describes - in a sense - our lack of knowledge about the universe
1 1. Preliminaries 2
U. Equivalence classes of the relation R, called granules, represent an elementary portion of knowledge we are able to perceive due to R. Using only the indiscernibility relation, in general, we are not able to observe individual objects from U but only the accessible granules of knowledge described by this relation. The concept of indiscernibility is central in rough set theory. Let I = (U; A) be an information system, where U is a non-empty nite set (the universe of discourse) and A is a non-empty nite set of attributes such that a : U Va for every a A. ! 2 Va is the set of values that attribute a may take. For any P A there is an asso- ciated equivalence relation R, also denoted by IND(P ), also called indiscernibility relation, given as:
R = IND(P ) = (x; y) U U : for all a P; a(x) = a(y) f 2 2 g If (x; y) IND(P ), then x and y are indiscernible by the attributes from P . The 2 partition of U, generated by R is denoted by U=R and the equivalence classes of the indiscernibility relation are denoted [x]R. Let X U. X can be approximated using only the information contained within R by constructing the lower and the upper approximations of X as below.
R(X) = x :[x]R X f g R(X) = x :[x]R X = f \ 6 ;g
Fig. 1.1. Rough set concept 1. Preliminaries 3
R is called the lower approximation operator and R is called the upper approxi- mation operator. The positive, negative and boundary regions of a subset X of U are de ned by: POSR(X) = R(X)
NEGR(X) = U R(X) BNDR(X) = R(X) R(X) The positive region contains all objects of U that can be classi ed to the equiva- lence classes of U with respect to the equivalence relation R . The boundary region,
BNDR(X), is the set of objects that can possibly, but not certainly, be classi ed in this way. The negative region, NEGR(X), is the set of objects that cannot be classi ed to classes of U=R. Now we formally state the de nition of a rough set.
De nition 1.1.1 A set X is called rough (inexact) with respect to R if and only if the boundary region of X is non-empty, that is, BNDR(X) = : Otherwise it will be 6 ; a crisp (exact) set.
In case of decision systems, the set X to be approximated may be de ned by the given decision attribute(s). For example, if d is a given decision attribute having the domain yes, no then X may be de ned as: f g X = x U : d(x) = yes f 2 g Example 1.1.2 [61] Suppose we want to describe the classi cation of basic trac signs to a novice. We start by saying that there are three main classes of trac signs corresponding to
Warning (W), Interdiction (I), Order (O). Then, we say that these classes may be distinguished by such attributes as the Shape (S) and the principal color (PC) of the sign. Finally, we give a few examples of trac signs, like those shown in Table 1.1. These are
(a) sharp right turn,
1 (b) speed limit of 50 km h ; 1. Preliminaries 4
(c) no parking,
(d) go ahead.
Table 1.1. Examples of trac signs described by S and PC
Here "Class" is the decision attribute and S and PC are the condition attributes. Let R be the indiscernibility relation induced by S and PC. Then the sets of signs indiscernible by S and PC, that is, R, are as follows:
[a]R = a ; [b]R = b ; [c]R = [d]R = c; d : f g f g f g The lower and the upper approximations of the classes W= a , I= b,c and f g f g O= d are as below: f g R(W ) = a R(W ) = a f g f g R(I) = b R(I) = b; c; d f g f g R(O) = R(O) = c; d ; f g These approximations can be interpretted as below:
Class W includes sign (a) certainly and possibly no other sign, Class I includes sign (b) certainly and possibly signs (b), (c) and (d), Class O includes no sign certainly and possibly signs (c) and (d). 1. Preliminaries 5
The terms certainly and possibly refer to the absence or presence of ambiguity between the description of signs by S and PC from the one side, and by "Class", from the other side.
One can de ne the following four basic classes of rough sets, that is, four categories of vagueness:
1. A set X is roughly R de nable, if and only if R(X) = and R(X) = U. 6 ; 6 2. A set X is internally R unde nable, if and only if R(X) = and R(X) = U. ; 6 3. A set X is externally R unde nable, if and only if R(X) = and R(X) = U. 6 ; 4. A set X is totally R unde nable, if and only if R(X) = and R(X) = U. ; The intuitive meaning of this classi cation is the following. A set X is roughly R de nable means that with respect to R we are able to decide for some elements of U that they belong to X and for some elements of U that they belong to Xc. A set X is internally R unde nable means that with respect to R we are able to decide for some elements of U that they belong to Xc, but we are unable to decide for any element of U whether it belongs to X. A set X is externally R unde nable means that with respect to R we are able to decide for some elements of U that they belong to X, but we are unable to decide for any element of U whether it belongs to Xc. A set X is totally R unde nable means that with respect to R we are unable to decide for any element of U whether it belongs to X or Xc. A rough set X can be also characterized numerically by the following coecient.
De nition 1.1.3 Let (U; R) be an approximation space. Then R(X) R(X) = j j R(X) is called the accuracy of approximation , where X denotes the cardinality of j j X = : 6 ;
Obviously, 0 R(X) 1: If R(X) = 1 then X is crisp with respect to R (X is precise with respect to R), and otherwise, if R(X) < 1;X is rough with respect to R (X is vague with respect to R). It can be seen that the lower approximation R(X) is greatest de nable set contained in X, and R(X) is the least de nable set containing X with respect to the equivalence relation R: The other properties of lower and upper approximations operators have been listed in the following theorem. 1. Preliminaries 6
Theorem 1.1.4 Let (U; R) be an approximation space. For any X;Y U; the fol- lowing properties hold for the lower and upper approximations.
1. R(X) X R(X); 2. X Y = R(X) R(Y ) and R(X) R(Y ); ) 3. R( ) = = R( ); ; ; ; 4. R(U) = U = R(U);
5. R(R(X)) = R(X) = R(R(X));
6. R(R(X)) = R(X) = R(R(X));
7. R(X Y ) = R(X) R(Y ); \ \ 8. R(X Y ) R(X) R(Y ); \ \ 9. R(X) R(Y ) R(X Y ); [ [ 10. R(X) R(Y ) = R(X Y ); [ [ 11. (R(X))c = R(Xc);
12. (R(X))c = R(Xc);
13. X is de nable R(X) = X R(X) = X R(X) = R(X); () () () 14. If X or Y are de nable, then R(X) R(Y ) = R(X Y ) and R(X Y ) = [ [ \ R(X) R(Y ): \ Properties 1, 2 and 5 of the above theorem show that R is an interior operator on U and is well-known that it induces a topology on U, that is, for any X U X if and only if R(X) = X 2 A topology de ned above is called the induced topology by R: The next two results have been taken from [29].
Proposition 1.1.5 Let R be an equivalence relation and be a topology induced by R. Then we have
R(X) = (R(Xc))c; that is; R is a closure operator on the topological space (U; ): 1. Preliminaries 7
The de nable sets can be characterized using topology as below.
Proposition 1.1.6 For every X U, the following conditions are equivalent. 1. X is de nable with respect to R;
2. X is an open subset in the topological space (U; ); where is the topology induced by R;
3. X is a closed subset in the topological space (U; ).
It follows from the above that the topological space (U; ) induced by an equivalence relation R on U has the property that for every subset X of U :
X is an open set if and only if it is a closed set.
1.1.2 Rough membership functions
In crisp set theory, characteristic function is employed to determine membership of an element to a given set subject to the available information. The interpretation of membership function in rough set theory is somehow di erent. It quanti es the degree of relative overlap between the set to be approximated and the equivalence class to which it belongs. In other words, it gives degree of belongingness of an element to a set subject to the given equivalence relation. Given an indiscernibilty relation R on U, the rough membership function R : U [0; 1] is de ned as follows X !
R [x]R X X (x) = j \ j [x]R j j where : denotes the cardinality. j j The rough membership function expresses conditional probability that x belongs to X given R and can be interpreted as a degree that x belongs to X in view of information about x expressed by R. The meaning of rough membership function can be depicted as shown in Fig. 1.2. 1. Preliminaries 8
[x]R
X
x
R X (x) = 0
[x]R
X
x
R 0 < X (x) < 1
[x]R
X
x
R X (x) = 1 Fig. 1.2. Membership of an element through rough membership function
The rough membership function can be used to de ne approximations and the boundary region of a set, as shown below:
R R(X) = x U : X (x) = 1 f 2 R g R(X) = x U : X (x) > 0 f 2 gR BNDR(X) = x U : 0 < (x) < 1 f 2 X g Theorem 1.1.7 For a given approximation space (U; R); the rough membership func- tion has the following properties:
1. R (x) = 1 if and only if x R(X) X 2 2. R (x) = 0 if and only if x U R(X) X 2 R 3. o < (x) < 1 if and only if x BNDR(X) X 2 R R 4. c (x) = 1 (x) for any x U X X 2 R R R 5. X Y (x) max(X (x); Y (x)) for any x U [ 2 R R R 6. X Y (x) min(X (x); Y (x)) for any x U: \ 2 1. Preliminaries 9
1.1.3 Generalizations of rough sets
The rough set theory (RST ) proved to be a useful tool in supporting data-related tasks such as classi cation, decision making and description. Many extensions of rough set theory have been proposed by generalizing the equivalence relation, using multiple relations and using the subset operator. The illustration in g. 1.3 shows the main RST extensions in relation to the aspects of the theory they extend. Some of them are discussed brie y in below.
Rough set theory
subset operator• mapping based single relation based multiple relation based extensions extensions extensions based extensions
Variable•precision T•rough sets Generalized Multi•granulation rough sets rough sets rough sets
Dominance based rough sets
Characteristic relation based rough sets
Nbhd rough sets
Fig. 1.3. A taxonomy of rough set extensions
Generalized rough sets [83]
Given a binary relation R and two elements x; y U; if xRy; we say that y is R related 2 to x: A binary relation may be more conveniently represented by a mapping r : U ! 2U : r(x) = y U : xRy f 2 g 1. Preliminaries 10
That is, r(x) consists of all R-related elements of x: Following Pawlak's approach, we de ne two unary set-theoretic operators apr and apr :
apr(A) = x : r(x) A f g = x U :for all y U; xRy implies y A f 2 2 2 g apr(A) = x : r(x) A = f \ 6 ;g = x U : there exists a y U such that xRy and y A f 2 2 2 g The set apr(A) consists of those elements whose R-related elements are all in A; and apr(A) consists of those elements such that at least one of whose R related elements is in A. The pair (apr(A); apr(A)) is referred to as the generalized rough set of A induced by R: Operators apr; apr : 2U 2U are referred to as the generalized rough ! set operators. The set r(x) may be interpreted as a neighnorhood of x. Hence, apr and apr are indeed the interior and closure of A: In case where R is an equivalence relation, generalized rough set operators reduce to the operators in Pawlak rough set model. For an arbitray relation, generalized rough set operators do not necessarily satisfy all the properties in Pawlak rough set models. Nevertheless, following properties hold in any rough set model, that is, independent of the properties of the binary relation.
apr(U) = U; apr( ) = ; ; ; A B = apr(A) apr(B); apr(A) apr(B); ) apr(A) = (apr(Ac))c; apr(A) = (apr(Ac))c; apr(A B) = apr(A) apr(B); \ \ apr(A B) apr(A) apr(B); [ [ apr(A B) = apr(A) apr(B); [ [ apr(A B) apr(A) apr(B): \ \ In order to construct a rough set model so that the above properties hold, it is necessary to impose certain conditions on the binary relation R. In fact, each of these properties corresponds to a property of the binary relation. A relation R is a serial relation if for all x U there exists y U such that 2 2 xRy: A relation is a re exive relation if for all x U the relationship xRx holds. A 2 1. Preliminaries 11 relation is symmetric relation if for all x; y U; xRy implies yRx holds. A relation 2 is transitive relation if for three elements x; y; z U; xRy and yRz imply xRz: A 2 relation is Euclidean when for all x; y; z U; xRy and xRz imply yRz: Since the 2 approximation operators are de ned through the mapping r; it is more convenient to express equivalently the conditions on a binary relation as follows: serial: for all x U; r(x) = : 2 6 ; re exive: for all x U; x r(x): 2 2 symmetric: for all x; y U; if x r(y); then y r(x): 2 2 2 transitive: for all x; y U; if y r(x); then r(y) r(x): 2 2 Euclidean: for all x; y U; if y r(x); then r(x) r(y): 2 2 The rough set models are named according to the properties of the binary relation. For example, a rough set model constructed from a symmetric relation is referred to as a symmetric rough set model. The ve properties of a binary relation, namely, the serial, re exive, symmetric, transitive, and Euclidean properties, include ve properties of the approximation op- erator, stated as below.
relation type property holds serial: apr(A) apr(A) re exive: apr(A) A symmetric: apr(A) apr(apr(A)) transitive: apr(A) apr(apr(A)) Euclidean: apr(A) apr(apr(A)) By combining these properties, one can construct more rough set models. For instance, if R is re exive and symmetric, that is, R is a compatibility (also called tolerance) relation, we obtain the rough set model built using a compatibility rela- tion. A compatibilty relation is a serial relation but not necessarily a transitive or an Euclidean relation.
T-rough sets [80]
Due to imprecise human knowledge, some times it is not possible to nd a suitable equivalence relation among the elements of the universe set U: Therefore, researchers felt the need to generalize the rough set theory. One of the generalizations was pre- sented by Yamak et al. 1. Preliminaries 12
De nition 1.1.8 Let X and Y be two non-empty universes. Let T be a set valued mapping given by T : X P (Y ): Then the triplet (X;Y;T ) is referred to as a ! generalized approximation space. Any set valued function from X to P (Y ) de nes a binary relation by setting T = (x; y): y T (x) : Obviously, if is an arbitrary f 2 g binary relation from X to Y, then it can de ne a set valued mapping T : X P (Y ) ! by T(x) = y Y :(x; y) where x X: For any set A Y a pair of lower and f 2 2 g 2 upper approximations T (A) and T (A) are de ned by
T (A) = x X : T (x) A f 2 g and T (A) = x X : T (x) A = : f 2 \ 6 ;g The pair (T (A); T (A)) is referred to as a generalized rough set, and T and T are referred to as lower and upper generalized approximation operators, respectively.
Multi-granulation rough sets
The Pawlak's rough set theory is based on a single granulation. Using more than one binary relations, Qian et al. [50] presented the concept of multi-granulation rough sets which is stated below.
De nition 1.1.9 Let R1; R2; :::; Rn be n independent equivalence relations over a universe set X and A X: The lower and upper multi-granulation rough approx- imations of A in X are de ned respectively as
n
( Ri)(A) = x X [x]Ri A for some i = 1; 2; :::; n i=1 f 2 j g P and n
( Ri)(A) = x X [x]Ri A = for all i = 1; 2; :::; n : i=1 f 2 j \ 6 ; g The boundaryP region of A X under MGRS environment is de ned as n n Bnd n (A) = ( Ri)(A) ( Ri)(A): Ri i=1 n i=1 i=1 P P P 1.2 Fuzzy sets
Traditional set theory is based on binary, or two-valued, logic. Given a universe set U, a subset A of U can be de ned in several ways: Suppose that U is the set of integers. The subset of prime numbers less than 10 can be speci ed by listing its members: A = 2; 3; 5; 7 or by providing de ning properties f g A = x : x is a prime number less than 10 : f g 1. Preliminaries 13
Alternately, we de ne a subset A by its characteristic function, which is also de- noted by the set name, A : U 0; 1 from U into the binary set 0; 1 given ! f g f g by 1 if x A A(x) = 2 ( 0 if x = A 2 Zadeh [86] de ned a fuzzy subset A of U as a function A : U [0; 1], that is, a ! characteristic function from U into the interval [0; 1]. The function from U to [0; 1] is called membership function of A, usually denoted by A: The value A(x) (or A(x)) is called the grade of membership of the point x in the fuzzy set A or the degree to which the point x belongs to the set A. For example, the fuzzy subset A of \the numbers close to 10" could be de ned by the membership function 1 (x) = A 1 + 5(x 10)2 Fig. 1.4. highlights the di erence between the concepts of membership in the crisp and fuzzy set theories.
a b
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0 0.5 1 0.5 1 universe of discourse universe of discourse Fig. 1.4. Membership function according to (a) crisp set theory (b) fuzzy set theory
1.2.1 Operations on fuzzy sets
The set of all fuzzy subsets of U will be denoted by (U). FP De nition 1.2.1 Let f; g (U). If f(x) g(x) for all x U, then f is said to 2 FP 2 be contained in g, and we write f g (or g f). 1. Preliminaries 14
Clearly, the inclusion relation is a partial order on (U). FP De nition 1.2.2 Let f; g (U): The union and intersection of fuzzy subsets f 2 FP and g of U for any x U is de ned respectively as follows: 2 (f g)(x) = f(x) g(x) [ _ (f g)(x) = f(x) g(x) \ ^ De nition 1.2.3 The complement of a fuzzy subset f (U) is denoted and de ned 2 FP as follows:
f 0(x) = 1 f(x); for all x U: 2 De nition 1.2.4 The fuzzy subsets of U, denoted by 0 and 1; de ned respectively as
0(x) = 0 1(x) = 1
for all x U; are called the empty (or null) fuzzy subset and the whole (or full) 2 fuzzy subset of U respectively.
1.3 Hesitant fuzzy sets
In this section we recall some basic notions related to the hesitant fuzzy sets. Xu et al. [78] de ned hesitant fuzzy sets in the following way:
De nition 1.3.1 Let U be a xed set. A hesitant fuzzy set (HFS) on U is given in terms of a function that when applied to U returns a subset of [0; 1] which can be represented as the following:
= (x; h (x)) : x U F f F 2 g where h (x) is a set of values in [0; 1], denoting the possible membership degrees F of the element x U to the set . For convenience, h (x) is called a hesitant fuzzy 2 F F element.
If h is a hesitant fuzzy set on U; then h(x) is a set of values in [0; 1] which represents the possible membership degrees of the element x U to the hesitant fuzzy set h: h(x) 2 will be called a hesitant fuzzy element (HFE) and its cardinality is called length of the HFE. Throughout, HF (U) will represent the collection of all hesitant fuzzy sets on U: 1. Preliminaries 15
Throughout, h(x) will be a nite subset for every hesitant fuzzy set h and any x U: 2 For a hesitant fuzzy element h(x); Torra [69] de ned its lower and upper bounds as below: lower bound: h (x) = min : h(x) ; f 2 g upper bound: h+(x) = max : h(x) : f 2 g Some related notions and operations on hesitant fuzzy elements by Yang et al. [81] are listed below.
De nition 1.3.2 Let x; y U and h1(x) and h2(y) be two hesitant fuzzy elements 2 on U: Then, their union ( ), intersection ( ) and containment ( ) are de ned in the Y Z following way:
1. Union: h1(x) Y h2(y) = h1(x) h2(y): max h1 (x); h2 (y) , f 2 [ f + g+ 2. Intersection: h1(x) Z h2(y) = h1(x) h2(y): min h1 (x); h2 (y) ; f 2 [ + f + g 3. Containment: h1(x) h2(y) h (x) h (y) and h (x) h (y): () 1 2 1 2 The above operations are de ned for hesitant fuzzy elements. For hesitant fuzzy sets, the following operations are de ned.
De nition 1.3.3 For hesitant fuzzy sets h, h1 and h2 on U; the following operations have been de ned:
1. Containment (Yang et al. [81]): h1 is contained in h2; denoted by h1 h2; if and only if h1(x) h2(x) for all x U; 2 2. Union (Torra [69]): union of h1 and h2 is the hesitant fuzzy set h1 dh2 de ned as (h1 h2)(x) = h1(x) h2(x) = h1(x) h2(x): max h (x); h (x) for all d Y f 2 [ f 1 2 g x U; 2 3. Intersection (Torra [69]): intersection of h1 and h2 is the hesitant fuzzy set + h1 h2 de ned as (h1 h2)(x) = h1(x) h2(x) = h1(x) h2(x): min h (x); e e Z f 2 [ f 1 h+(x) for all x U; 2 g 2 4. Complement (Torra [69]): complement of h is the hesitant fuzzy set hc which is de ned as (hc)(x) = 1 for all x U; [h(x)f g 2 2 5. Di erence (Liao [35]): di erence of h1 and h2 is the hesitant fuzzy set h1 h2 de ned as (h1 h2)(x) = ; where 1 h1(x); 2 h2(x)f g 2 S 2 1 2 if and = 1 1 2 1 2 2 = 6 ( 0 otherwise
Remark 1.3.4 1. The operations e; d are de ned on hesitant fuzzy sets while opera- tions Z; Y are de ned on the corresponding hesitant fuzzy elements. We shall use the symbol c both to denote complements of hesitant fuzzy sets and hesitant fuzzy elements. 1. Preliminaries 16
2. Yang et al. in [81] proved that for hesitant fuzzy sets h1; h2 HF (U); h1 h2 2 e h1; h2 and h1; h2 h1 h2. d 3. Gang et al. proved in [20] that hesitant fuzzy union and intersection can also be expressed in the following form:
(h1 d h2)(x) = max 1; 2 ; 1 h1(x)[; 2 h2(x) f g 2 2 (h1 e h2)(x) = min 1; 2 : 1 h1(x)[; 2 h2(x) f g 2 2
4. By induction, for an arbitrary nite collection hi (i I = 1; 2; :::; n ) of hesitant 2 f g fuzzy sets, their union and intersection will be taken as follows:
( hi) = (((h1 h2) h3)::: hn) idI d d d 2 ( hi)(x) = (((h1 h2) h3)::: hn) ieI e e e 2 5. Equality of hesitant fuzzy sets will be taken as follows:
Two hesitant fuzzy sets h1 and h2 are equal if h1(x) = h2(x) for all x U: 2 It should be noted that the equality h1(x) = h2(y) for any x; y U is not followed by 2 h1(x) h2(y) and h2(y) h1(x): For example, if h1(x) = 0:3; 0:4; 0:7 and h2(y) = f g 0:3; 0:5; 0:6; 0:7 then h1(x) h2(y) and h2(y) h1(x): Nevertheless, h1(x) = h2(y): f g 6 Yang et al. [81] investigated the following operational laws for hesitant fuzzy elements.
Theorem 1.3.5 For any hesitant fuzzy elements h1(x); h2(y) and h3(z) on U, the following properties hold: 1. Idempotent: h1(x) Z h1(x) = h1(x); h1(x) Y h1(x) = h1(x); 2. Commutativity: h1(x) Z h2(y) = h2(y) Z h1(x); h1(x) Y h2(y) = h2(y) Y h1(x); 3. Associativity: h1(x) Z (h2(y) Z h3(z)) = (h1(x) Z h2(y)) Z h3(z); h1(x) Y (h2(y) Y h3(z)) = (h1(x) Y h2(y)) Y h3(z); 4. Distributivity: h1(x) Z (h2(y) Y h3(z)) = (h1(x) Z h2(y)) Y (h1(x) Z h3(z)); h1(x) Y (h2(y) Z h3(z)) = (h1(x) Y h2(y)) Z (h1(x) Y h3(z)); c c c c c 5. De Morgan's laws: (h1(x) Z h2(y)) = h1(x) Y h2(y); (h1(x) Y h2(y)) = h1(x) Z c h2(y); c c 6. Double negation: (h1(x)) = h1(x):
Torra [69] and Yang et al. [81] mentioned some special hesitant fuzzy sets on U.
1. empty set: h(x) = 0 for all x U; f g 2 2. full set: h(x) = 1 for all x in U; f g 1. Preliminaries 17
3. complete ignorance: for any x U; h(x) = [0; 1]; 2 4. set for a nonsense x : h(x) = for some x U; ; 2 5. constant set : h(x) = A for all x U; where A is some xed subset of [0; 1]: 2
Throughout, the empty hesitant fuzzy set will be denoted by h0 while full hesitant fuzzy set will be denoted by hU : To compare hesitant fuzzy elements, Xia and Xu [77] introduced score function. They made the following assumptions for this function: -The values of all the HFEs are arranged in an increasing order. -The HFEs have the same length when they are compared. Therefore, if any two HFEs have di erent lengths, the shorter one will be extended by adding the maximum element until both HFEs have the same length.
De nition 1.3.6 (Xia and Xu [77])For a hesitant fuzzy element h(x), score func- tion is de ned as
s(h(x)) = #h(x) 0 1 h(x) 2X where s( ) is the score function and #@h(x) is theA number of elements in h(x):
Example 1.3.7 Let h1 = 0:3; 0:5; 0:8 and h2 = 0:1; 0:4; 0:9 be two hesitant fuzzy f g f g elements. Their union, intersection, complements and di erence have been calculated in below: h1 h2 = 0:3; 0:5; 0:8 0:1; 0:4; 0:9 Y f g Y f g = 0:3; 0:4; 0:5; 0:8; 0:9 f g h1 h2 = 0:3; 0:5; 0:8 0:1; 0:4; 0:9 Z f g Z f g = 0:1; 0:3; 0:4; 0:5; 0:8 f g hc = 0:2; 0:5; 0:7 ; hc = 0:1; 0:6; 0:9 1 f g 2 f g h1 h2 = 0:22; 0:44; 0:17; 0:78; 0:67; 0:0 f g Also, the scores of h1 and h2 are caluclated in below:
0:3+0:5+0:8 s(h1) = 3 = 0:53 0:1+0:4+0:9 s(h2) = 3 = 0:47
Here score of h1 is greater than the score of h2. Chapter 2
Fuzzi ed Rough Sets
2.1 Introduction
There may be some data set in which values of attributes are both symbolic and numerical. In these cases, it will be dicult to nd the underlying patterns from the classical rough set approach. If the attribute values in the given information system are not numerical, that is, they are qualitative rather than quantitative then it may be necessary to fuzzify the given universe set or the subset to be approximated. Also, in some situations clustering of elements subject to the given attributes may not be useful or even may not be possible. A slight di erence in the attribute value may lead two elements in di erent classes which, in real, may be very close to each other. Then, rather than considering elements' indistinguishability, assessing their similarity (to a certain degree) can be more suitable. Taking into account the diculty in handling the above situations, many authors introduced hybridization of fuzzy and rough set theories in di erent ways [39, 40, 41, 53]. From application point of view, constructive approach proves to be more useful which was initiated by Dubois and Prade [17, 18]. In this approach, lower and upper approximations are constructed by means of fuzzy relations. In the above stated paper, t-norm min and its dual conorm max were used to de ne fuzzy similarity relation. This approach was further generalized by Radzikowska and Kerre [52], by using an implication operator and a t-norm. They de ned a broad family of fuzzy rough sets with respect to a fuzzy similarity relation which they called ( ; ) fuzzy rough sets. This approach was further generalized I T by Ouyang et al. [43] in which they used transitivity instead of classical sup-min T transitivity of the fuzzy relation. These approaches of de ning fuzzy rough set use similarity (or similarity) relations in which the classes are disjoint. Less e ort T has been made to explore the structure of these fuzzy rough sets using intransitive fuzzy relations. Also, interpretation of the results depend highly on the particular
18 2. Fuzzi ed Rough Sets 19 fuzzy relation being used but the choice of appropriate fuzzy relation has not been addressed in them. This issue is addressed and resolved in this chapter. Some other generalizations can be seen in [3, 8, 10, 12, 43, 51, 54, 67, 68, 72, 76, 84]. In the present chapter, a new technique for fuzzi cation of rough set is introduced which involves both the fuzzi cation of the information system if the attribute val- ues are linguistic terms and a tolerance (intransitive) fuzzy relation which is used to measure the compatibility in indiscernible objects; the objects which don't have exactly the same attributes but they are similar or compatible up to a certain degree . Upper and lower approximations are then de ned using the above concepts for approximating subsets of the given universe set. The technique is further extended by using implications and t-norms. For application purposes, two algorithms have been introduced for approximating subsets of the universe based on their similarity among objects, subject to the given attributes.
2.2 Fuzzi ed rough approximations for indiscernible objects
Many attempts have been made to fuzzify rough approximations. Two approaches have mainly been adopted; rst, to fuzzify the set to be approximated so that the objects meet its characteristics to varying degrees and second, to highlight similarity between objects through a fuzzy relation and categorizing the objects into granules with soft boundaries based on their similarity to one another. In this section, we have developed an algorithm for approximating subsets of the universe by combining both of these extended approaches. The algorithm is exible enough to handle both the numerical as well as linguistic attribute values in the infor- mation table. In some situations, there are some positive as well as negative attributes. Then the same attribute value may have di erent meaning for the two attributes. For example, the value "high" has exactly the opposite meaning regarding the attributes "hardworking" and "carelessness". Thus, to fuzzify the attribute values; the algorithm gives a margin to use separate fuzzy membership functions for positive and negative attributes. Also, since we are considering the objects which are all discernible, that is, no two objects have exactly the same attributes, categorizing the objects into classes with sharp boundaries will be of no use. Therefore, a fuzzy relation has been de ned to emphasize the relationship between the objects.
De nition 2.2.1 A fuzzy subset f (U U) is called a fuzzy binary relation or 2 F simply a fuzzy relation on U: A fuzzy relation f is called a serial fuzzy relation if for each x U; there exists y U such that f(x; y) = 1; f is referred to as a re exive 2 2 2. Fuzzi ed Rough Sets 20 fuzzy relation if f(x; x) = 1 for all x U; f is referred to as a symmetric fuzzy 2 relation if f(x; y) = f(y; x) for all x; y U; f is referred to as a transitive fuzzy 2 relation if f(x; z) (f(x; y) f(y; z)) for all x; y; z U. y _U ^ 2 A re exive, symmetric2 fuzzy relation on U is called proximity relation or com- patibility relation in U (also called tolerance relation): These proximity relations can intuitively be interpreted as measures of `likeness' or `sameness' among the elements. When R is a fuzzy compatibility relation, compatibility classes are de ned in terms of a speci ed membership degree : An compatibility class is a subset A of U such that R(x; y) for all x; y A: 2 A fuzzy binary relation is called a similarity relation (or fuzzy equivalence re- lation) if it is a re exive, symmetric and transitive fuzzy relation. The notion of similarity relation is a natural generalization of an equivalence relation.
De nition 2.2.2 Let (U; R) be a fuzzy approximation space, where U is a nite (non- empty) set of objects and R is a fuzzy tolerance relation characterized by its membership function R : U U [0; 1]: For any (0; 1]; the fuzzi ed lower and upper rough ! 2 approximations for a given set X U are de ned as c R (X) = x U : R(x; y) < for all y X f 2 2 g and
R (X) = x U : R(x; y) for some y X : f 2 2 g The pair (R (X); R (X)) is referred to as fuzzi ed rough set. Positive, negative and boundary regions of X U for any (0; 1] are denoted and de ned as: 2
POSR(X) = R (X); c NEGR(X) = (R (X)) ;
BNDR(X) = R (X) R (X): Based on the given information table, the elements in the positive, negative and boundary region can be interpreted, respectively as certain members, certain non- members and possible members (may or may not be) of X up to a relational degree of (0; 1]: Throughout, R will be considered as tolerance fuzzy relation until otherwise 2 speci ed.
Example 2.2.3 Let U = x1; x2; x3; x4; x5; x6 be a universe set with fuzzy tolerance f g relation R as de ned in Table 2.1.
The lower and upper fuzzi ed approximations of a subset X = x1; x3; x5 with f g = 0:9 are calculated as below: 2. Fuzzi ed Rough Sets 21
R x1 x2 x3 x4 x5 x6 x1 1 0.5 0.9 0.6 0.2 0.3 x2 0.5 1 0.7 0.1 0.4 0.5 x3 0.9 0.7 1 0.3 0.7 0.8 x4 0.6 0.1 0.3 1 0.9 0.6 x5 0.2 0.4 0.7 0.9 1 0.8 x6 0.3 0.5 0.8 0.6 0.8 1
Table 2.1: A fuzzy tolerance relation
c Here X = x2; x4; x6 . Thus, R (X) = x1; x3 as R(x1; y) < 0:9 for all f g 0:9 f g y x2; x4; x6 and R(x3; y) < 0:9 for all y x2; x4; x6 : 2 f g 2 f g Also, as R(x1; x1) = 1 > 0:9;
R(x3; x3) = 1 > 0:9;
R(x4; x5) = 0:9;
R(x5; x5) = 1 > 0:9;
so, R0:9(X) = x1; x3; x4; x5 . Thus, ( x1; x3 ; x1; x3; x4; x5 ) is 0:9 fuzzi ed f g f g f g rough set. Also, POSR(X) = x1; x3 ;NEGR(X) = x2; x6 and BNDR(X) = f g f g x4; x5 : f g Remark 2.2.4 In the above de nition, signi es the level or degree up to which the relationship or compatibility among the objects is to be considered so that we may interpret lower approximation as a subset of U containing only those elements which are not related (up to to ) to any element outside X, that is, Xc while the upper approximation contains those elements of U which are related (up to ) to any element of X:
De nition 2.2.5 For a given (0; 1] and a fuzzy relation R in U characterized by 2 the membership function R : U U [0; 1]; the objects x and y in U will be called ! indiscernible if R(x; y) :
Proposition 2.2.6 Let (U; R) be a fuzzy approximation space and 1; 2 (0; 1] be 2 such that 1 2:Then R (X) R (X) and R (X) R (X): 1 2 2 1 c Proof. For any x R 1 (X); R(x; y) < 1 for all y X : But since 1 2;we 2 c 2 have R(x; y) < 2 for all y X : Thus, x R (X) showing that R (X) R (X): 2 2 2 1 2 Similarly, if x R (X); then R(x; y) 2 1 for some y X depicting that 2 2 2 x R (X): Hence, R (X) R (X): 2 1 2 1 2. Fuzzi ed Rough Sets 22 2.3 Proposed Algorithm
Taking (U; R) as fuzzy approximation space, A as set of attributes and T; the infor- mation table, following algorithm is proposed to approximate subsets of U: As already mentioned, it is assumed that no two objects have the same attribute values and the attribute values may be qualitative or quantitative.
Algorithm 2.3.1 Step 1: Input the triplet (U; A; T ): Step 2: If the attribute values in the information table are de ned by linguistic terms instead of numerical values; assign a fuzzy membership function to the attribute values. Corresponding to each attribute value, choose the point where the membership is maximum. If the attribute values are numerical, move to step 3. Step 3: For each pair of objects x; y U; nd fuzzy relation R in U U by the 2 membership function n
R(x; y) = R(xi; yi)=n (1) i=1 X where n is the number of attributes and for any ith attribute (i = 1; 2; :::n); xi; yi are the corresponding fuzzy values of the objects x and y; respectively. Here, R is a fuzzy relation in [0; 1] [0; 1] for the fuzzi ed attribute values characterized by the membership function
(xi; yi) = 1 xi yi for all i = 1; 2; :::; n: (2) R j j Step 4: Input the set of objects X U to be approximated. Step 5: Input degree of relationship (0; 1]: 2 Step 6: Calculate lower and upper approximations of X as de ned in De nition 2.2.2.
Step 2 assigns grades to attribute values of the decision matrix in Step 1 from
[0; 1]. For a pair of fuzzi ed attribute values (xi; yi) for an ith attribute corresponding to the pair of objects (x; y), the fuzzy relation R is a similarity relation which gives the degree to which both the values are related, that is, it measures similarity between them while the relation R averages these degrees for each pair of objects. Step 3 calculates this relation R. In steps 4 and 5, we input (crisp) set to be approximated (which may be a decision partition class) and the degree to which the elements are to be kept similar. The last step calculates approximations to signify the certain and probable members. 2. Fuzzi ed Rough Sets 23
Fuzzification Finding similarity Finding lower and of data set relation upper approximations
Fig. 2.1. Algorithm for nding approximations using alpha-fuzzi ed rough sets
Remark 2.3.2 R; being a re exive and symmetric relation, is a fuzzy compatibility relation which measures compatibility among the elements. Intuitively, compatibility or proximity relations are used to model the vague concept of `likeness' or `sameness'. Ovchinnikov [44] clari ed the idea behind proximity relation as follows. Suppose X is a nite set of objects and A is a nite set of their attributes such that each object a X has at least one attribute p A. Let X(p) be the set of all a X with attribute 2 2 2 p. We say that an object a resembles an object b if and only if they belong to the same subset X(p) for some attribute p A. This idea can be formalized as follows. Let R 2 be a binary relation on X de ned by aRb if there is p A such that a; b X(p). Then 2 2 R is a re exive symmetric binary relation. But, when there are three or more values of attribute p; fuzzy binary relations are indispensable to give relation among objects. Therefore, the need of fuzzy relation R in Algorithm 2.3.1 to nd the resemblance or compatibility among the objects subject to the given attributes is apparent.
2.3.1 Illustrative Example
Given a set of cost contingency techniques U = u; v; w; x; y; z = pre-determined f g f percentage, expert judgement, risk analysis, regression analysis, Delphi technique and PERT, NASA's joint con dence level model ; a set of attributes A = complexity, g f team e ort, accuracy, risk concerned and an information table given below,we are g aiming to nd the approximations of a subset X = u; v; x of U using Algorithm f g 2.3.1. Step 1: Input the triplet (U; A; T ) given by Table 2:2: Step 2: The semantics of the given linguistic terms (qualitative attribute values) 2. Fuzzi ed Rough Sets 24
u v w x y z complexity low low medium high medium high team e ort low low medium low high low accuracy low medium high medium high medium risk concerned low medium high medium high high
Table 2.2: Information Table Attribute/Objects u v w x y z complexity 0.75 0.75 0.5 0.25 0.5 0.25 team e ort 0.25 0.25 0.5 0.25 0.75 0.25 accuracy 0.25 0.5 0.75 0.5 0.75 0.5 risk concerned 0.75 0.5 0.25 0.5 0.25 0.25
Table 2.3: Fuzzi ed Information Table are given by triangular fuzzy membership functions as in g. 1 and g. 2 of g. 2.2.
low medium high high medium low
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
Fig. 1. Semantics of three terms for Fig. 2. Semantics of three terms for the the set {team effort, accuracy} set {complexity, risk concerned}
Fig. 2.2. Fuzzy membership functions
We notice that regarding the attributes "complexity" and "risk concerned", "low" is the best value while "high" is the worst value. In the rest of the attributes "low" and "high" have exactly the opposite meaning. Thus, we have assigned two di erent membership functions for the linguistic term sets complexity, risk concerned and f g team e ort, accuracy : The points where these terms have their maximum member- f g ship value are shown in Table 2.3. Step 3: Evaluate fuzzy relation as de ned in Step-3 of Algorithm 2.3.1. The resulting fuzzy relation R in U U is given in Table 2:4. Step 4: Input the set X = u; v; x of objects which is to be approximated. f g Step 5: Input degree of relationship = 0:9375: Step 6: Calculate lower and upper approximations of X = u; v; x using De ni- f g tion 2.2.2 as below: Here Xc = w; y; z . From Table 2.4 it can be seen that u and v are the only f g elements whose degree of relationship with respect to R to all the elements in Xc, 2. Fuzzi ed Rough Sets 25
R u v w x y z u 1 0.875 0.625 0.75 0.5625 0.6875 v 0.875 1 0.75 0.875 0.6875 0.8125 w 0.625 0.75 1 0.75 0.9375 0.8125 x 0.75 0.875 0.75 1 0.6875 0.9375 y 0.5625 0.6875 0.9375 0.6875 1 0.75 z 0.6875 0.8125 0.8125 0.9375 0.75 1
Table 2.4: Fuzzy relation among objects (contingency techniques) that is, w; y; z is less than = 0:9375: Therefore, R (X) = u; v : While the upper f g approximation contains all those elements whose degree of relationship with at least one element of X is 0:9375 or more. Since R is re exive fuzzy relation, elements in X will be in its upper approximation. Among the rest of objects (w; y; z); z is the only object such that R(z; x) = 0:9375 = for the element x in X and so is in R (X). It is clear from the above table that w and y are related to all elements of X with degree less than 0:9375. Thus, R (X) = u; v; x; z : f g We conclude that the techniques u (pre-determined percentage) and v (expert judgement) can be classi ed as members of X with respect to R up to a classi cation degree = 0:9375: While x (regression analysis) and z (NASA's joint con dence level model) can be classi ed as possible members of X with respect to R up to a classi cation degree = 0:9375:
Remark 2.3.3 The purpose of introducing degree of relationship is to give a margin to the data sets in which no two objects has exactly the same attribute values. Instead of taking them to be exactly similar; we are considering the second strongest degree to which two elements might be related. If it still does not work, we'll choose the third strongest degree and so on. Therefore, in Example 2.3.1 the degree is chosen to be 0.9375 which is the second strongest degree to which two elements can be related.
2.4 Properties of fuzzi ed rough approximations
Lemma 2.4.1 For any x; y U; the membership function R(x; y) of fuzzy relation 2 R de ned in equation 1 satis es the following property:
R(x; y) = 1 if and only if x = y: 2. Fuzzi ed Rough Sets 26
Proof. If x = y for any x; y U; then 2 n R(x; x) = R(xi; xi)=n i=1 n P = (1 xi xi )=n i=1 j j n = P(1)=n = n=n = 1: i=1 P Conversely, if R(x; y) = 1 then
n R(xi; yi)=n = 1 i=1 n P = (1 xi yi ) = n ) i=1 j j n P = n xi yi = n ) i=1 j j n P = xi yi = 0 ) i=1 j j = Pxi yi = 0 for all i=1,2,...,n. ) j j = xi = yi for all i=1,2,...,n. ) = x = y (Since no two objects have the same attributes). )
Following properties of the fuzzi ed lower and upper rough approximations hold.
Theorem 2.4.2 Let (U; R) be fuzzy approximation space, (0; 1]: Then for X U 2 we have
(1) R (X) X R (X);
(2) R ( ) = = R ( ); ; ; ;
(3) R (U) = U = R (U);
c c (4) R (X ) = (R (X)) ;
c c (5) R (X ) = (R (X)) ;
(6) X Y = R (X) R (Y ) and R (X) R (Y ); )
(7) R1 R2 = R2 (X) R1 (X) and R1 (X) R2 (X); ) (8) R (X Y ) = R (X) R (Y ); \ \ (9) R (X Y ) R (X) R (Y ); [ [ 2. Fuzzi ed Rough Sets 27
(10) R (X Y ) = R (X) R (Y ); [ [
(11) R (X Y ) R (X) R (Y ): \ \ Proof. (1) (3) Straightforward. (4) For any x U; 2 c x R (X ) 2 c c R(x; y) < for all y (X ) = X () 2 R(x; y) for any y X () 6 2 c x R (X) x (R (X)) : () 62 () 2 (5) For any x U; 2 c x R (X ) 2 c R(x; y) for some y X () 2c R(x; y) < for all y X () 6 2 x R (X) x (R (X))c: () 62 () 2 c c c (6) For any x R (X); R(x; y) < for all y X : In particular, since Y X ; 2 c 2 we may write R(x; y) < for all y Y which implies that x R (Y ): Thus, 2 2 R (X) R (Y ): Also, if x R (X); then R(x; y) for some y X Y which implies that 2 2 x R (Y ): Hence, R (X) R (Y ): 2 (7) The proof follows from De nition 2.2.2 and the fact that R1 R2 implies R (x; y) R (x; y) for all x; y U: 1 2 2 (8) By using part (6) above and the fact that X Y X;Y we may write \ R (X Y ) R (X);R (Y ) \ and so, R (X Y ) R (X) R (Y ): \ \
For the reverse inclusion, let x R (X) R (Y ): Then, x R (X) and x R (Y ), 2 c \ 2 c 2 that is, R(x; y) < for all y X and R(x; z) < for all z Y which gives c c2 c 2 R(x; u) < for all u X Y = (X Y ) and so x R (X Y ): Thus, R (X) 2 [ \ 2 \ \ R (Y ) R (X Y ) and hence R (X) R (Y ) = R (X Y ): \ \ \ (9) Since X;Y X Y , by part (6) of this theorem, we may write [ R (X);R (Y ) R (X Y ) [ and so, R (X) R (Y ) R (X Y ): [ [ (10) As X;Y X Y; by using part (6) we get [
R (X); R (Y ) R (X Y ) [ and so, R (X) R (Y ) R (X Y ): [ [ 2. Fuzzi ed Rough Sets 28
For the reverse inclusion, let x R (X Y ): Then 2 [
R(x; y) for some y X Y 2 [ The above condition can be expressed as
R(x; y) for some y X 2 or R(x; y) for some y Y 2
This gives us that x R (X) or x R (Y ), that is, x R (X) R (Y ): Hence, 2 2 2 [ R (X Y ) = R (X) R (Y ): [ [ (11) From part (6) above and the fact that X Y X;Y we get \
R (X Y ) R (X); R (Y ) \ and so, R (X Y ) R (X) R (Y ): \ \
Remark 2.4.3 (1) If = 1 then R (X) = X = R (X). Indeed, from part 1 of the
above theorem we have R (X) X R (X) for any X U and for any x X; c 2 from Lemma 2.4.1 we have R(x; z) = 1 for all z X , that is, R(x; z) < 6 2 (= 1) for all z Xc which implies X R (X) and so X = R (X): Also, for 2 any x R (X); R(x; y) (= 1) for some y X: But since R(x; y) [0; 1] 2 2 2 we get R(x; y) = 1 which implies, from Lemma 2.4.1, that x = y: Thus, x X 2 as y X which implies R (X) X: Hence, R (X) = X: 2 (2) The reverse inclusions in parts (8) and (10) do not hold which can be justi ed by Example 2.3.1. If we take X = u; v; w and Y = v; w; x and = 0:625: f g f g Then, by De nition 2.2.2, we get R (X) = ;R (Y ) = and since X Y = f g f g [ u; v; w; x ; we have R (X Y ) = u : Thus, R (X Y ) R (X) R (Y ): Also, f g [ f g [ 6 [ R (X) = u; v; w; x; y; z and R (Y ) = u; v; w; x; y; z . And since X Y = f g f g \ v; w , we have R (X Y ) = v; w; y; z which shows that R (X) R (Y ) f g \ f g \ 6 R (X Y ): \ c c (3) Since U = R (X) (R (X) R (X)) R (X); and the three sets (R (X) ; (R (X) [ [ R (X));R (X) )are disjoint, they form partition of U. These three sets are the negative, boundary and positive regions respectively.
2.5 Fuzzi ed rough membership function
In crisp set theory, an element either belongs to or does not belong to a set. Thus membership of an element to a crisp set can be represented through a characteristic 2. Fuzzi ed Rough Sets 29 function which maps an element to 1 if it belongs to the set and to 0 if it does not belong to that set. In rough set theory, the notion of membership function is generalized and is interpreted as a degree of certainty to which an element belongs to the set. Following the same concept, fuzzi ed rough membership function can be de ned. De nition 2.5.1 For a given fuzzy approximation space (U; R) and (0; 1]; we 2 de ne a relation in U U as below:
xy if and only if R(x; y) :
Then class for any x U will be [x] = y U : xy : 2 f 2 g This relation is a compatibility relation and each class is a compatibility class. Using this relation we can de ne fuzzi ed rough membership function [x] X R : U [0; 1] for a crisp set X asR (x) = j \ j for all x U: X X [x] ! j j 2 Example 2.5.2 In Example 2.2.3, for = 0:9; classes are given below: [x1] = x1; x3 = [x3]; [x2] = x2 , [x4] = x4; x5 = [x5]; [x6] = x6 : f g f g f g f g Using these classes, the membership of elements of U with X will be as below:
R R X (x1) = x1; x3 = x1; x3 = 1 = X (x3); R jf gj jf gj X (x2) = = x2 = 0; R j;j jf gj R X (x4) = x5 = x4; x5 = 1=2 = X (x5); R jf gj jf gj (x6) = = x6 = 0: X j;j jf gj Remark 2.5.3 fuzzi ed rough membership function and classes can be used to de ne the approximations as below:
R (1) R (X) = x U : (x) = 1 = x U :[x] X : Indeed, f 2 X g f 2 g
x R (X) 2 c R(x; y) < for all y X () 2 x y for any y Xc () 6 2 [x] X [x] X = [x] () () \ R (x) = 1: () X
R (2) R (X) = x U : (x) > 0 = x U :[x] X = ? : Indeed, f 2 X g f 2 \ 6 g
x R (X) 2 R(x; y) for some y X () 2 xy for some y X () 2 [x] X = ? [x] X > 0 () \ 6 () j \ j R (x) > 0: () X 2. Fuzzi ed Rough Sets 30
R (3) Clearly, BNDR(X) = R (X) R (X) = x U : 0 < (x) < 1 : f 2 X g Proposition 2.5.4 The fuzzi ed rough membership function has the following prop- erties.
(1) R (x) = 1 if and only if x R (X); X 2
R c (2) (x) = 0 if and only if x (R (X)) ; X 2
R (3) 0 < (x) < 1 if and only if x BR (X); X 2
R R (4) c (x) = 1 (x) for any x U; X X 2
R R R (5) X Y (x) max(X (x); Y (x)) for any x U; [ 2 R R R (6) X Y (x) min(X (x); Y (x)) for any x U: \ 2 Proof. Straightforward.
2.6 Measures associated with fuzzi ed rough sets This section presents some measures associated with fuzzi ed rough sets. The main concern is to measure completeness of knowledge provided by a given tolerance fuzzy relation. Pawlak [47] presented the accuracy and roughness measures associated with rough set approximations. The accuracy measure is the ratio of the lower approximation to the upper approximation while the roughness measure is complement of the accuracy measure. The purpose of introducing this measure is to capture the degree of com- pleteness of our knowledge about the set X or to express quality of an approximation. The roughness measure as complement of accuracy measure is interpreted as degree of incompleteness. As a generalization of these measures, we introduce an accuracy measure using fuzzi ed rough approximations as follows. De nition 2.6.1 Let (U; R) be a fuzzy approximation space and (0; 1]. The 2 accuracy measure of a subset X of U by R is de ned as follows.
R (X) (X;R) = j j A R (X)
where X = and X denotes the cardinality of the set X. 6 ; j j
Theorem 2.6.2 Let U be a universe set and R;R1;R2 be fuzzy relations over U. For
= X U; the accuracy measure (X;R) satis es the following properties. ; 6 A 2. Fuzzi ed Rough Sets 31
(1) (X;R) = 1 if and only if R (X) = R (X); A
(2) (X;R) = 0 if and only if R (X) = ; A ;
(3) For a xed R (X); (X;R) strictly monotonically increases with R (X) ; A j j
(4) For a xed R (X) = ; (X;R) strictly monotonically decreases with R (X) ; 6 ; A j j
(5) If 1 2; then (X;R) (X;R); A 1 A 2
(6) If R1 R2; then (X;R2) (X;R1): A A Proof. Straightforward.
Proposition 2.6.3 If R is an equivalence relation, then 1 fuzzi ed rough approxi- mations degenerate into rough approximations.
Proof. Straightforward. Gediga and D•untsch [21] proposed a measure for the precision of approximation of X U for a given partition; which is not e ected by the approximation of X: For a given equivalence relation R, it is denoted and de ned as below.
apr (X) R (X;R) =
X j j In fuzzi ed rough environment it can be extended as follows. R (X) (X;R) = j j X j j Clearly, (X;R) (X;R): It may be noted that this measure requires complete A knowledge of X; whereas does not. It can be interpreted as the relative number of A elements in X which can be approximated by R.
Proposition 2.6.4 If 1 2; then (X;R) (X;R): 1 2 Proof. Straightforward. Yao [82] pointed out that the term `accuracy' should be de ned and interpreted accurately as it might be misleading in some cases. He revised some of the properties of the accuracy measure given by Pawlak [46, 47] and presented another measure of the `completeness' of knowledge, or the accuracy of approximations. The suggested measure is as follows.
POSR(X) + NEGR(X) (X;R) = j j Uj j apr (X) +j (japr (X))c R j R j = j j U apr (X) +j aprj (Xc) R R = j j Uj j j j 2. Fuzzi ed Rough Sets 32
Following his approach, we can extend this measure using fuzzy approximation space as below.
De nition 2.6.5 Let (U; R) be a fuzzy approximation space and (0; 1]. The 2 measure of quality of approximation of a subset X of U by R is de ned as follows.
R (X) + R (Xc) (X;R) = j j j j U j j The corresponding roughness measure is
(X;R) = 1 (X;R) R (X) R (X) = j U j j j
The measure can be interpreted as the measure of quality of approximation of the X generated partition X;Xc by R. f g
Theorem 2.6.6 For a given fuzzy approximation space (U; R) and X U; (X;R) satis es the following properties.
(1) (X;R) = 0 if and only if (R (X) = ; R (X) = U), ;
(2) For a xed R (X)(may be empty or non-empty); (X;R) strictly monotonically decreases with R (X) :
Proof. Straightforward.
The measure can serve well to check the quality of the X generated partition only, that is, X;Xc . To deal with decision partitions that contains more than two f g decision classes, we have the generalized measure of quality of approximation as follows.
R (Y ) : Y U=D (D;R) = fj j 2 g U P j j
Proposition 2.6.7 If D1 D2 (that is, for any D1(x) D1; there exists D2(x) D2 2 2 such that D1(x) D2(x)); where Di represents the partition induced by the decision b b b b b b attribute Di, then (D1;R) (D2;R): b b b Proof. Straightforward.
Example 2.6.8 Consider a fuzzy relation R over a universe set U = x1; x2; x3; f x4 as de ned in Table 2.5. Taking = 0:7; the lower approximations, the upper g approximations and the values of measures ; ; for all non-empty subsets X of A U have been shown in Table 2.6. 2. Fuzzi ed Rough Sets 33
R x1 x2 x3 x4 x1 1 0.3 0.9 0.6 x2 0.3 1 0.4 0.7 x3 0.9 0.4 1 0.5 x4 0.6 0.7 0.5 1
Table 2.5: A tolerance fuzzy relation
X R (X) R0:7(X) 0:7 A x1 x1; x3 0 0 1/2 f g ; f g x2 x2; x4 0 0 1/2 f g ; f g x3 x1; x3 0 0 1/2 f g ; f g x4 x2; x4 0 0 1/2 f g ; f g x1; x2 x1; x2; x3; x4 0 0 0 f g ; f g x1; x3 x1; x3 x1; x3 1 1 1 f g f g f g x1; x4 x1; x2; x3; x4 0 0 0 f g ; f g x2; x3 x1; x2; x3; x4 0 0 0 f g ; f g x2; x4 x2; x4 x2; x4 1 1 1 f g f g f g x3; x4 x1; x2; x3; x4 0 0 0 f g ; f g x2; x3; x4 x2; x4 x1; x2; x3; x4 1/2 2/3 1/2 f g f g f g x1; x2; x3 x1; x3 x1; x2; x3; x4 1/2 2/3 1/2 f g f g f g x1; x2; x4 x2; x4 x1; x2; x3; x4 1/2 2/3 1/2 f g f g f g x1; x3; x4 x1; x3 x1; x2; x3; x4 1/2 2/3 1/2 f g f g f g x1; x2; x3; x4 x1; x2; x3; x4 x1; x2; x3; x4 1 1 1 f g f g f g Table 2.6: All non-empty subsets of U with their accuracy measures 2. Fuzzi ed Rough Sets 34
t-norms 1) standard min operator M (x; y) = min x; y ; (the largest t-norm [28]) T f g 2) algebraic product P (x; y) = xy; 3) the bold intersectionT (also calledLukasiewicz t-norm) L(x; y) = max 0; x + y 1 ; T f g t-conorms 1) standard max operator M (x; y) = max x; y ; S f g 2) the probabilistic sum P (x; y) = x + y xy; S 3) bounded sum L(x; y) = min 1; x + y : S f g Table 2.7: Some famous t-norms and t-conorm
2.7 Generalization of fuzzi ed rough sets using logical connectives
In this section, we shall de ne an extension of fuzzi ed rough sets by using some of the logical connectives, particularly, implication and triangular norms. We'll start this section by recalling some basic de nitions of logical connectives. A triangular norm, or shortly t-norm, [27] is an increasing, associative and commutative mapping : [0; 1] [0; 1] [0; 1] satisfying (1; x) = x for all x [0; 1]: T ! T 2 A triangular conorm, or shortly t-conorm, is any increasing, commutative and associative mapping : [0; 1] [0; 1] [0; 1] satisfying (0; x) = x for all x [0; 1]: S ! S 2 A negation operator is a decreasing mapping : [0; 1] [0; 1] satisfying N ! (0) = 1 and (1) = 0: Standard negation operator (or a negator) denoted by s N N N is de ned as s(x) = 1 x for all x [0; 1]:A negator is called involutive if N 2 N ( (x)) = x for all x [0; 1]: N N 2 An implication operator (or implicator) is any mapping : [0; 1] [0; 1] I ! [0; 1] satisfying (1; 0) = 0 and (1; 1) = (0; 1) = (0; 0) = 1: It is called border I I I I implication if it satis es the condition that (1; x) = x for all x [0; 1]: I 2 Based on a t-norm ; t-conorm and an involutive negator ; an implicator T S N I de ned as (x; y) = ( (x); y) for all x; y [0; 1] is called an S implicator. I S N 2 Another implicator based on a t-norm de ned as (x; y) = sup [0; 1] : I T I f 2 (x; y) ; provided that is continuos, is called an R-implicator (residual im- T g T plicator). Some famous t-norms, t-conorms, negations and implicators have been mentioned in Tables 2.7 and 2.8. For further details about implication operators, we refer to [55].
Theorem 2.7.1 [52] Every S implicator and R implicator is a border implicator. Proof. Straightforward. 2. Fuzzi ed Rough Sets 35
S implicators 1)Lukasiewicz implicator L(x; y) = min 1; 1 x + y I f g based on L and s; S N 2) Kleene-Dienes implicator KD(x; y) = max 1 x; y I f g based on M and s; S N 3) Kleene-Dienes-Lukasiewicz implicator P (x; y) = 1 x + xy IS based on P and s: S N R implicators 1) theLukasiewicz implicator L based on L; I T 2) the G•odelimplicator G(x; y) = 1 for x y and I G(x; y) = y elsewhere, based on M ; 3)I the Gaines implicator (x; y)T = 1 for x y and I4 (x; y) = y=x elsewhere, based on P : I4 T Table 2.8: Some famous implicators
By using Remark 2.5.3, we can re-write lower and upper fuzzi ed rough ap- proximations as
R (X) = x U : ( for all y U) xy = y X f 2 2 ) 2 g (3) and R (X) = x U : ( there exist) xy y X f 2 ^ 2 g where = and are an implicator and a triangular norm, which have been used to ) ^ interpret the conditions [x] X and [x] X = ? respectively. \ 6 Radzikowska and Kerre [52] introduced a broad class of fuzzy rough sets which they called ( ; ) fuzzy rough sets, when represents implicator and ; a triangular I T I T norm. The de nition is stated below.
De nition 2.7.2 Let FAS = (X;R) be a fuzzy approximation space and and be I T a border implicator and a t-norm, respectively. An ( ; ) fuzzy rough approximation ; I T in FAS is a mapping AprI T : (X) (X) (X) de ned for every A (X) FAS F ! F F 2 F by ; AprI T (A) = (F (A); F T (A)); FAS AS AS I where FAS (A) and FAST (A)) are de ned for every x X as: I 2
FAS (A)(x) = inf (R(x; y);A(y)); y XI I 2 FAST (A)(x) = sup (R(x; y);A(y)): y XT 2
The fuzzy set FAS (A) (respectively FAST (A)) is called an lower (respectively I I upper) fuzzy rough approximation of A in FAS: T 2. Fuzzi ed Rough Sets 36
From equations 3, it is clear that the lower and upper fuzzi ed rough approxi- mations mean:
(x) = 1 if and only if for all y U (R(x; y) = X (y) = 1) R (X) 2 ) (4) (x) = 1 if and only if for all y U (R(x; y) X (y) = 1) R (X) 2 ^ As pointed out by Radzikowska and Kerre [52], = and can be interpreted as ) ^ an implicator and a t-norm ; respectively. With this interpretation, we get the I T following extended form of fuzzi ed rough approximations in De nition 2.2.2. De nition 2.7.3 Let (U) be the collection of all fuzzy sets over the nite universe F U; R : U U [0; 1] be the fuzzy relation and (0; 1]: Let a fuzzy relation ! 2 R : U U [0; 1] be de ned as ! R(x; y) if R(x; y) R (x; y) = ( 0 if R(x; y) <
Then lower and upper ( ; ) fuzzi ed rough approximations are fuzzy sets de- I T ned for any f (U) as 2 F
R (f)(x) = inf (R (x; y); f(y)) y UI 2 and R (f)(x) = sup (R (x; y); f(y)) y UT 2 where is a border implicator and is a t-norm. The pair (R (f); R (f)) will be I T called ( ; ) fuzzi ed rough set. I T
Example 2.7.4 Consider the fuzzy relation R over a universe set U = x1; x2; x3; f x4 as de ned in Table 2.5 in Example 2.6.8 and a fuzzy set f de ned below. g
0.6 if x = x1 8 0.9 if x = x2 f(x) = > > 0.2 if x = x3 <> 0.5 if x = x4 > Using De nition 2.7.3, and taking> = 0:5; R is calculated as in Table 2.9. :> Using the same de nition and taking G and M as the implicator and the T-norm I T (de ned in Tables 2.7 and 2.8), the lower and upper approximations of the fuzzy set f as calculated in below.
R (f)(x1) = inf G(R0:5(x1; x1); f(x1)); G(R0:5(x1; x2); f(x2)); 0:5 fI I G(R0:5(x1; x3); f(x3)); G(R0:5(x1; x4); f(x4)) I I g = inf G(1; 0:6); G(0; 0:9); G(0:9; 0:2); G(0:6; 0:5) fI I I I g = inf 0:6; 1; 0:2; 0:5 = 0:2: f g 2. Fuzzi ed Rough Sets 37
R0:5 x1 x2 x3 x4 x1 1 0 0.9 0.6 x2 0 1 0 0.7 x3 0.9 0 1 0.5 x4 0.6 0.7 0.5 1
Table 2.9: Fuzzy relation R-0.5
R0:5(f)(x1) = sup M (R0:5(x1; x1); f(x1)); M (R0:5(x1; x2); f(x2)); fT T M (R0:5(x1; x3); f(x3)); M (R0:5(x1; x4); f(x4)) T T g = sup M (1; 0:6); M (0; 0:9); M (0:9; 0:2); M (0:6; 0:5) fT T T T g = sup 0:6; 0; 0:2; 0:5 = 0:6: f g Similarly, the rest of the calculations can be made that are summarized as below.
0.2 if x = x1 8 0.5 if x = x2 R0:5(f)(x) = > > 0.2 if x = x3 <> 0.2 if x = x4 > > :> 0.6 if x = x1 8 0.9 if x = x2 R0:5(f)(x) = > > 0.6 if x = x3 <> 0.7 if x = x4 > > 2.7.1 Proposed Algorithm :
The relation R has been used to ful l the requirement R(x; y) as in equation 4. Thus, two elements will be related if their degree of relationship is at least : Using these ( ; ) approximations, Algorithm 2.3.1 can be revised as follows. I T Algorithm 2.7.5 Step 1: Input the triplet (U; A; T ) where U is a nite universe, A is the set of attributes and T is the given information table. Step 2: Fuzzify the information table by assigning suitable values from the interval [0; 1] to each attribute value. Step 3: For each pair of objects x; y U; nd fuzzy relation R in U U by the 2 membership function n
R(x; y) = R(xi; yi)=n i=1 X where n is the number of attributes and for any ith attribute (i = 1; 2; :::n); xi; yi are the corresponding fuzzy values of the objects x and y; respectively. Here, R is 2. Fuzzi ed Rough Sets 38
R u v w x y z u 1 0.75 0 0.5 0 0 v 0.75 1 0.5 0.75 0 0.625 w 0 0.5 1 0.5 0.875 0.625 x 0.5 0.75 0.5 1 0 0.875 y 0 0 0.875 0 1 0.5 z 0 0.625 0.625 0.875 0.5 1
Table 2.10: A fuzzy relation R-alpha a fuzzy relation in [0; 1] [0; 1] for the fuzzi ed attribute values characterized by the membership function
(xi; yi) = 1 xi yi for all i = 1; 2; :::; n: R j j Step 4: Input degree of relationship (0; 1]: 2 Step 5: Find R : U U [0; 1] as in De nition 2.7.3. ! Step 6: Input a fuzzy set f over U which is to be approximated.
Step 7: Find ( ; ) lower and upper fuzzi ed rough approximations as in Def- I T inition 2.7.3 by using a suitable border implication and a t-norm : I T The above algorithm presents an expedient technique for approximations of fuzzy sets. The upper and lower approximations are fuzzy sets which can be determined by choosing a suitable implicator and a t-norm.
Example 2.7.6 Consider the information system in Example 2.3.1. Steps 1-3 are the same as in algorithm 2.3. Taking = 0:5; we nd R using table 2.4. The calculated values are shown in table 2.10. Suppose we want to approximate the fuzzy set f on U de ned below.
0.7 if s = u 8 0.4 if s = v > 0.3 if s = w > f(s) = > > 0.8 if s = x <> 0.1 if s = y > > 0.5 if s = z > > Using De nition 2.7.3, the Lukasiewicz> implicator L and the Lukasiewicz T-norm : I L (de ned in Tables 2.7 and 2.8), we nd the lower and upper 0:5 fuzzi ed approxi- T 2. Fuzzi ed Rough Sets 39 mations as below. 0.65 if s = u 8 0.4 if s = v > 0.225 if s = w > R (f)(s) = > > 0.625 if s = x <> 0.1 if s = y > > 0.5 if s = z > > :> 0.7 if s = u 8 0.55 if s = v > 0.30 if s = w > R (f)(s) = > > 0.80 if s = x <> 0.175 if s = y > > 0.675 if s = z > > > 2.8 Comparative analysis:
Rough set theory introduced by Pawlak [46] was based on crisp equivalence classes. It works well in discrete data sets. But de ning classes with sharp boundaries on data sets containing continuos values is impractical and sometimes impossible. This fact led to the development of fuzzy rough set theories. Pioneering work in this regard can be seen in [17]. Other noteworthy techniques include a generalized approach by Radzikowska and Kerre [53] and a further generalization of this work by Ouyang et al. [43]. Dubois and Prade [17] suggested two almost similar approaches to combine fuzzy and rough sets. First one is to use fuzzy similarity relation in place of crisp equivalence relation and in the second approach; fuzzy partition is introduced to get fuzzy granules. This approach proved its worth in de ning granules of objects with soft bound- aries and thus making rough set theory more practical. Radzikowska and Kerre [53] presented another approach (( ; )-fuzzy rough sets) of combining the two theories I T by using fuzzy implicator and triangular norm. They categorized fuzzy rough sets in three classes based on S-, R- and QL- implicators and proved that fuzzy rough set by Dubois and Prade [17] is a particular case of the newly presented fuzzy rough sets. The above mentioned techniques of combining fuzzy and rough sets were based on fuzzy similarity relation that is analogous to equivalence relation in crisp case. These approaches are valid only if transitivity in fuzzy relation is considered as sup-min transitivity. In case of -transitivity (in which standard min operator is replaced by T -norms), these approaches are unsuccessful in producing disjoint clusters which are T expected in case of similarity relations. 2. Fuzzi ed Rough Sets 40
Fuzzy rough sets for -similarity relations (also called tolerance relations) were T discussed by Ouyang et al. [43] as a generalization of ( ; )-fuzzy rough sets. They I T proved some properties that were not satis ed by ( ; )-fuzzy rough sets. I T These hybrid models produced worthy approaches in combining fuzzy and rough sets but are de cient in two main aspects. Firstly, they do not provide a way to cluster objects by using fuzzy relations that are not (sup-min or ) transitive. Secondly, they T yield fuzzy sets as approximations whose interpretations are highly dependent on the choice of fuzzy relation. The induction of di erent fuzzy relations will highly e ect the membership values in lower and upper approximations as well as their interpretations.
For example, R1(x; y) = min(x; y) and R2(x; y) = 1 x y are both fuzzy relations j j but they di er a lot in their actual meanings. The former one gives minimum possible values while the latter gives nearness among the elements. Also, the same membership value may have di erent meanings with respect to di erent fuzzy relations. Like, in the above mentioned relations, R1(0:2; 0:4) = R2(0:1; 0:9) = 0:2 whose interpretation is totally di erent with respect to R1 and R2. Due to these reasons the existing theories are dicult to implement in practical situations. D'eer et al. [16] established a mathematical framework for the most relevant fuzzy rough set models proposed in the literature and critically evaluated them. They in- vestigated and compared these models on the basis of their properties. They declared three main properties as the most important ones among the rest and discussed them in detail. These are stated in parts 1, 6 and 7 of Theorem 2.4.2. It is concluded in the said paper that more likely none of the existing models is able to retain all the proper- ties of Pawlak's original model. Fuzzy variable precision rough set model presented by Zhao et al. [87] satis es most theoretical properties, yet it does not satisfy the inclu- sion property, that is, property 1 of Theorem 2.4.2. The other fuzzy rough set models that satisfy some of these properties, but not all, are proposed in [11, 13, 57, 58, 84]. In this chapter, we presented a technique to overcome these shortcomings. Approx- imations have been constructed with a view point of clustering objects with respect to fuzzy relations that may not be transitive, that is, compatibility relations. A control parameter has been introduced which allows to choose relational degree among the elements as close as we wish to without being exactly similar. Fuzzy relation R used in the presented algorithms measures compatibility among the elements. This fuzzy relation R and crisp lower and upper approximations make it easy to interpret and implement the results in practical situations. The approximation operators proposed in this paper satisfy all the properties marked as important in [16]. The inclusion prop- erty (Theorem 2.4.2, part 1) is intuitively expected to hold in any such structure. The set monotonicity property (Theorem 2.4.2, part 6) makes this technique possible and reliable to apply in classi cation problems. Furthermore, the relation monotonicity 2. Fuzzi ed Rough Sets 41
C1 C2 C3 C4 C5 C6 x1 0.8 0.1 0.1 0.5 0.2 0.3 x2 0.3 0.5 0.2 0.8 0.1 0.1 x3 0.2 0.2 0.6 0.7 0.3 0.2 x4 0.6 0.3 0.1 0.2 0.5 0.3 x5 0.3 0.4 0.3 0.3 0.6 0.1 x6 0.2 0.3 0.5 0.3 0.5 0.2 x7 0.3 0.3 0.4 0.2 0.6 0.2 x8 0.3 0.4 0.3 0.1 0.4 0.5 x9 0.3 0.2 0.5 0.4 0.4 0.2
Table 2.11: Samples of credit card evaluation problem property (Theorem 2.4.2, part 7) plays its role in applications like attribute reduction [16]. Thus, the presented model is more advantageous practically in the sense that it has the above mentioned signi cant properties. As a numerical experiment, we consider an evaluation problem of credit card ap- plicants formulated by Tsang et al. [71] (Example 3.1, Table 1) and approximate two (decision partition) sets by Dubois and Prade [17] technique and by algorithm 2.3.1.
The problem consists of nine elements (applicants) x1; x2; x3; x4; x5; x6; x7; x8; x9 and the applicants are described by six fuzzy attributes C1(best education);C2(better education);C3(good education);C4(high salary);C5(middle salary);C6(low salary). The information system is shown in Table 2.11. The sets to be approximated are
A = x1; x2; x4; x7 and B = x3; x5; x6; x8; x9 : The approximations by the approach f g f g in [17] are calculated using two fuzzy relations R1 and R2 de ned below.
R1(xi; xj) = min(Ck(xi);Ck(xj)) k
R2(xi; xj) = min 1 Ck(xi) Ck(xj) k f j jg The resulting lower and upper approximations are given in Table 2.12. The ap- proximations have also been calculated using algorithm 2.7.5 (taking M as t-norm T and KD as implicator) and the last column of this table shows approximations by I algorithm 2.3.1. The abrupt change in approximations (by changing fuzzy relation) can be noticed in C 1 and C 2 columns of the table. Depending upon R1 and R2, membership values of the approximations have di erent meanings and cannot be interpreted in the same manner. Since R2 is comparatively a better fuzzy relation to measure similarity among elements, we compare our results of algorithm 2.3.1 and algorithm 2.7.5 with C 2. FRS lower approximation of A contains three elements x1; x2 and x4 and it is evident from C 2 and C 3 of the table that membership of these three 2. Fuzzi ed Rough Sets 42
C 1 C 2 C 3 C 4 R1 R2 algo2( = 0:9) algo1( = 0:9) 0.9 x = x1 0.5 x = x1 1 x = x1 0.9 x = x2 0.4 x = x2 1 x = x2 8 8 8 AR > 0.9 x = x4 > 0.3 x = x4 > 1 x = x4 x1; x2; x4 > > > f g <> 0.8 x = x7 <> 0.1 x = x7 <> 0.07 x = x7 0 otherwise 0 otherwise 0 otherwise > > > > 0.2 x = x3 > 0.6 x = x3 > 0 x = x3 :> :> :> 0.1 x = x5 0.9 x = x5 0.93 x = x5 8 8 8 0.2 x = x6 0.9 x = x6 0.93 x = x6 x1; x2; x4; AR > > > f > 0.1 x = x8 > 0.7 x = x8 > 0 x = x8 x5; x7 > > > g < 0.2 x = x9 < 0.8 x = x9 < 0.9 x = x9 > 1 otherwise > 1 otherwise > 1 otherwise > > > > 0.8 x = x3 > > 1 x = x3 :> :> 0.4 x = x3 :> 0.9 x = x5 0.07 x = x5 8 0.1 x = x5 8 > 0.8 x = x6 8 > 0.07 x = x6 BR > 0.3 x = x8 > x3; x8; x9 > 0.9 x = x8 > > 1 x = x8 f g > > 0.1 x = x > < 0.8 x = x < 9 < 0.10 x = x 9 0 otherwise 9 > 0 otherwise > > 0 otherwise > > > > :> 0.5 x = x1 > :> 0.1 x = x1 :> 0 x = x1 0.6 x = x2 0.1 x = x2 8 0 x = x2 8 0.7 x = x4 8 x3; x5; x6; BR 0.1 x = x4 > 0 x = x4 f > > 0.9 x = x > x ; x ; x > 0.2 x = x > 6 > 0.93 x = x 7 8 9 < 7 < 0.9 x = x < 7 g 1 otherwise 7 1 otherwise > > 1 otherwise > > > > :> > :> Table 2.12::> Comparison of approximations 2. Fuzzi ed Rough Sets 43 elements (to lower approximation of A) is greater than rest of the elements. FRS upper approximation of A contains ve elements that have maximum (1) or 0:9 (2nd maximum) degree in C 2 and C 3: Similar interpretation holds for upper approximation of B: The elements in the lower approximation of B in C 4 have maximum or second maximum degree in C 3 (algorithm 2.7.5) but di er slightly in C 2. This di erence can be removed by appropriate choice of fuzzy relation in C 2. Although researchers have obtained remarkable achievements regarding fuzzy rough sets, yet there are a series of problems to be solved such as how fuzzy relation will be selected and how to interpret the results according to that relation. Without a satisfactory answer to these questions, the implementation of the existing fuzzy rough sets approaches to practical issues is questionable. In view of the above comparative results, the salient features of the proposed techniques can be listed as below; (1) The technique is compared with the state-of-the-art fuzzy rough set presented by Dubois and Prade [17] that uses a fuzzy similarity relation. The approximations by algorithm 1 and 2 use a fuzzy tolerance relation, yet there is no signi cant di erence among the results. (2) A given concept can be approximated either by two crisp sets (lower and upper approximations) using algorithm 1 or by fuzzy sets using algorithm 2 according to the situation. (3) The explicit tolerance fuzzy relation used makes the techniques easy to imple- ment and interpret. (4) The proposed approximation operators satisfy the important properties that are necessary from the application perspective. Thus, it is hoped that the algorithms presented in this paper will be more e ective in handling issues like risk analysis, classi cation problems, attribute reduction and pattern recognition.
2.9 Conclusion
We introduced a general approach for fuzzi cation of rough sets through a tolerance fuzzy relation and also investigated some properties of the proposed technique and its membership function. An algorithm has been presented to show practicality of this new interesting approach. Generalization of this technique using logical connectives accompanied by an algorithm is also discussed. Comparison of this technique with other renowned ones shows its validity and ecacy. The proposed algorithms in this chapter dealt with a special fuzzy proximity (in- transitive) relation. For an arbitrary fuzzy relation R; the approach presented in this chapter needs further investigations. Chapter 3
Generalized Hesitant Fuzzy Rough Sets
3.1 Introduction
As a generalization of fuzzy rough sets, the concept of generalized hesitant fuzzy rough sets (GHF RS) is presented in this chapter. It is an endeavor to de ne rough approx- imations of a collection of hesitant fuzzy sets over a given universe. To this end, elements of the universe are initially clustered using a set valued map and then hes- itant fuzzy sets are aggregated by using lower and upper approximation operators. These operators produce hesitant fuzzy sets which aggregate hesitant fuzzy elements. Structural and topological properties associated with GHF RS have also been exam- ined. The new model is further employed to design a three-way decision analysis technique which preserves many properties of classical techniques but needs less e ort and computation. Unlike the existing approaches, the alternatives can be clustered and selected jointly by using a set valued mapping. This feature makes its application area broader. Moreover, this method is applied to an example, where risk analysis issue is discussed for the selection of energy projects.
3.2 Approximations of Generalized hesitant fuzzy rough sets (GHF RS)
This section introduces a new technique for combining rough and hesitant fuzzy sets which involves no hesitant fuzzy relation. Its structural and topological properties have also been examined. In order to keep the things simple throughout this chapter, we shall consider the case of discrete structures only.
44 3. Generalized Hesitant Fuzzy Rough Sets 45
Several aggregation operators and measures have been de ned for hesitant fuzzy sets [74, 77, 78]. The main purpose in de ning these operators and measures is to average them or to nd distance or similarity between them. Instead of averaging hes- itant fuzzy elements, one may be interested in nding their minimum and maximum. For example, an investment company may be interested in nding maximum chances of pro t or minimum chances of loss for investing a sum of money in a project subject to some given factors (which may be taken as hesitant fuzzy sets). Also, most of these existing measures are based on algebraic sum and algebraic product of hesitant fuzzy elements. In practical applications, like in decision making problems, these measures will change the actual values provided by experts. This may be misleading in some situations. To overcome these problems, we introduce two hesitant fuzzy operators which will be called lower and upper hesitant fuzzy rough approximation operators. These operators are based on a given collection of hesitant fuzzy sets over a universe set U and a set-valued map T which maps each element of U to a non-empty subset of U: This map generalizes the concept of equivalence classes in rough set theory and allows us to choose clusters of elements without any restriction. Moreover, the oper- ators of hesitant fuzzy union Y and hesitant fuzzy intersection Z have been used to de ne the approximations so that the actual values will be retained.
3.2.1 (GHF RS) approximation operators and its properties
De nition 3.2.1 Let = hi : i I be a ( nite) collection of hesitant fuzzy sets H f 2 g on U; that is, HF (U) and T : U P (U) be a set valued mapping, where H ! P (U) = P (U) . The lower and upper approximations of with respect to T are n; H hesitant fuzzy sets T and T de ned for any x U as: H H 2
T (x) = Z Z hi x0 H i I x0 T (x) 2 2 and T (x) = Y Y hi x0 H i I x0 T (x) 2 2 respectively. The pair T ; T is called a generalized hesitant fuzzy rough set H H (GHFRS) with respect to T . is called hesitant fuzzy de nable if T = T . H H H Example 3.2.2 A student Z is planning to appear in a competitive exam. He has to choose three subjects among a list of optional subjects S1, S2, S3, S4, S5, S6 keeping in view their three attributes h1 (overall scoring), h2 (time for preparation) and h3
(length of syllabus). There is one more restriction. If he chooses subject S2 then he must choose S3 as well. He seeks advice of three experts and ask them to grade the subjects according to their attributes. These grades are given in Table 3.1 below: 3. Generalized Hesitant Fuzzy Rough Sets 46
h1 h2 h3 S1 (:2;:4;:3) :3;:5;:6 :6;:6;:7 f g f g S2 :8;:5;:6 :2;:4;:5 :7;:7;:5 f g f g f g S3 :1;:2;:5 :8;:7;:6 :4;:3;:7 f g f g f g S4 :4;:2;:3 :9;:6;:7 :6;:4;:3 f g f g f g S5 :6;:5;:8 :4;:4;:5 :3;:5;:7 f g f g f g S6 :3;:5;:5 :8;:5;:5 :3;:5;:6 f g f g f g Table 3.1: Hesitant fuzzy sets showing grades of student Z
Here = h1; h2; h3 is the collection of hesitant fuzzy sets and U = S1;S2;S3; H f g f S4;S5;S6 is the set of alternatives. According to the given restriction, we de ne the g mapping T : U P (U) as below: !
Si if i 1; 4; 5; 6 T (Si) = f g 2 f g ( S2;S3 if i 2; 3 f g 2 f g
According to De nition 3.2.1, the lower and upper approximations of the subjects S1 to S6 are as below:
T (S1) = :2;:4;:3 ; T (S1) = :6;:7 ; H f g H f g T (S2) = :2;:4;:5 ; T (S2) = :5;:6;:7;:8 ; H f g H f g T (S3) = :1;:2;:3;:4;:5 ; T (S3) = :8;:7;:6 ; H f g H f g T (S4) = :4;:2;:3 ; T (S4) = :9;:6;:7 ; H f g H f g T (S5) = :4;:5 ; T (S5) = :6;:5;:8;:7 ; H f g H f g T (S6) = :3;:5 ; T (S6) = :8;:5;:6 : H f g H f g The lower approximations give the minimum (expected) grading and upper approxima- tions give the maximum (expected) grading. By comparing these values, he will choose
S4, S5 and S6 as they have relatively better grading than the rest.
Example 3.2.3 Suppose that a software company desires to hire a system analysis engineer. After preliminary screening, four candidates y1, y2; y3 and y4 quali ed for further evaluation. A committee of three decision-makers, has been formed to select the suitable candidate for the post according to the following ve criteria:
(1) emotional steadiness (h1),
(2) oral communication skill (h2),
(3) personality (h3),
(4) past experience (h4),
(5) self-con dence (h5). The decision makers provide their preferences in anonymity and the decision matrix is given in Table 3.2. 3. Generalized Hesitant Fuzzy Rough Sets 47
h1 h2 h3 h4 y1 (:2;:4;:7) (:3;:5;:6) (:6;:6;:7) (:5;:6;:8) y2 (:2;:4;:5) (:5;:7;:7) (:1;:2;:5) (:6;:7;:8) y3 (:3;:4;:7) (:4;:4;:6) (:6;:7;:7) (:4;:6;:9) y4 (:5;:6;:8) (:4;:4;:5) (:3;:5;:5) (:5;:5;:8)
Table 3.2: Hesitant fuzzy decision matrix
Here = h1; h2; h3; h4 is the collection of hesitant fuzzy sets and U = y1; y2; H f g f y3; y4 is the set of alternatives. To categorize the objects (candidates) into classes g according to the preferences given in Table 3.2; we de ne the image sets of T : U ! P (U) as below.
T (x) = y U : s(hi(y)) s(hi(x)) < 0:05 for all i = 1; 2; 3; 4 f 2 j j g where s(h(x)) denotes the score of h(x) de ned in De nition 1.3.6. Thus, the image sets of mapping T : U P (U) will be as follows: ! yi if i 2; 4 T (yi) = f g 2 f g ( y1; y3 if i 1; 3 f g 2 f g According to De nition 3.2.1, the lower and upper approximations of the candidates y1 to y4 are as below:
T (y1) = T (y3) = :2;:3;:4;; 5;:6 ; T (y1) = T (y3) = :6;:7;:8;:9 ; H H f g H H f g T (y2) = :1;:2;:4;:5 ; T (y2) = :6;:7;:8 ; H f g H f g T (y4) = :3;:4;:5 ; T (y4) = :5;:6;:8 : H f g H f g The lower approximations give the minimum (expected) grading and upper approxi- mations give the maximum (expected) grading. By comparing these values, y1 and y3 have relatively the better grading than the rest.
Lemma 3.2.4 For any collection hi(xj): i I; j J of hesitant fuzzy elements f 2 2 g on U; Z hi (xj) Y hi (xj) : i I;j J i I;j J 2 2 2 2 Proof. For the given collection of hesitant fuzzy elements, we have