Dominance Relation and Rules in an

Incomplete Ordered Information System

Ming-Wen Shao,* Wen-Xiu Zhang †
Institute for Information and System Sciences, Faculty of Science,
Xi’an Jiaotong University, Xi’an, Shaan’xi 710049, P. R. China

Rough sets theory has proved to be a useful mathematical tool for classification and prediction.
However, as many real-world problems deal with ordering objects instead of classifying objects,
one of the extensions of the classical rough sets approach is the dominance-based rough sets
approach, which is mainly based on substitution of the indiscernibility relation by a dominance
relation. In this article, we present a dominance-based rough sets approach to reasoning in incom-
plete ordered information systems. The approach shows how to find decision rules directly from
an incomplete ordered decision table. We propose a reduction of knowledge that eliminates only
that information that is not essential from the point of view of the ordering of objects or decision
rules. © 2005 Wiley Periodicals, Inc.

1. INTRODUCTION

Rough sets theory, introduced by Pawlak,1 has been conceived as a tool to

conceptualize and analyze various types of data. It has important applications to
artificial intelligence and cognitive sciences, as a tool for dealing with vagueness
and uncertainty of facts, and in classification.2–8
The original rough sets theory approach does not consider attributes with
preference-ordered domains, that is, criteria. However, in many real situations, we
are often faced with the problems in which the ordering of properties of the con-
sidered attributes plays a crucial role. One such type of problem is the ordering of
objects. For this reason, Greco, Matarazzo, and Slowinski 9–13 proposed an exten-
sion of rough sets theory, called the dominance-based rough sets approach (DRSA)
to take into account the ordering properties of criteria. This innovation is mainly
based on substitution of the indiscernibility relation by a dominance relation. In
DRSA, where condition attributes are criteria and classes are preference ordered,
the knowledge approximated is a collection of upward and downward unions of

*Author to whom all correspondence should be addressed: e-mail: shaomingwen1837@

163.com.
†
e-mail: [email protected].

INTERNATIONAL JOURNAL OF INTELLIGENT SYSTEMS, VOL. 20, 13–27 (2005)

© 2005 Wiley Periodicals, Inc. Published online in Wiley InterScience
(www.interscience.wiley.com). • DOI 10.1002/int.20051
14 SHAO AND ZHANG
classes and the granules of knowledge are sets of objects defined using a domi-
nance relation. In recent years, many studies have been made in DRSA.14–17
According to whether or not there are missing data (null values), the data
table can be classified into two categories: complete and incomplete. Some impor-
tant results have recently been obtained for incomplete information systems 4,6,18–21
by utilizing the classical rough sets theory approach.
The problem of classification and rules extraction from ordered incomplete
information systems was also discussed in the context of the DRSA.22 In this arti-
cle, we study the problem of making decision rules from an incomplete ordered
decision table (DT) based on the use of the dominance relation. The proposed
approach is a extension of the original approach of a classical incomplete DT.
They maintain the good characteristics of the original one and boil down to the
original one when dominance relation is substituted by equivalence relation. The
extension also has a distinctive feature in comparison with the extension proposed
by Greco, Matarazzo, and Slowinski.22
As we know, the discernibility matrix 23 of an information system depicts com-
pletely the identification capability of the system, and therefore, all reducts of the
system are hidden in the discernibility function induced by the discernibility matrix.
For acquiring all reducts in an ordered information system (complete or incom-
plete) and extracting decision rules from a ordered decision table (complete or
incomplete), the technique of the discernibility function derived from the earlier
notion of a discernibility matrix is sufficient. Therefore, many theoretical methods
for finding all reducts in an ordered information system and extracting decision
rules from a ordered decision table have been developed around the construction
of a suitable and simple discernibility function equivalent to the original one
obtained directly from the system. And it is the focus of analysis in this article.

2. INCOMPLETE ORDERED INFORMATION SYSTEMS

An information system (IS) is an quadruple I ⫽ ~U, AT,V, f !, where U is a
finite nonempty set of objects and AT is a finite nonempty set of attributes, V ⫽
艛a僆AT Va and Va is a domain of attribute a, f : U ⫻ AT r V is a total function
such that f ~ x, a! 僆 Va for every a 僆 AT, x 僆 U, called an information function.
A decision table is a special case of an information system in which, among
the attributes, we distinguish one called a decision attribute. The other attributes
are called condition attributes. Therefore, I ⫽ ~U, AT 艛 $d %,V, f ! and AT 艚 $d % ⫽
⭋, where set AT contains so-called condition attributes and d, the decision attribute.
If the domain (scale) of a condition attribute is ordered according to a decreas-
ing or increasing preference, then the attribute is a criterion.

Definition 1. An information system is called an ordered information system

(OIS) if all condition attributes are criterions.

It is assumed that the domain of a criterion a 僆 AT is completely pre-

ordered by an outranking relation Ɒa ; x Ɒa y means that x is at least as good as
 DOMINANCE RELATION AND RULES 15
(outranks) y with respect to criterion a. In the following, without any loss of gen-
erality, we consider a condition criterion having a numerical domain, that is, Va 債 R
(R denotes the set of real numbers) and being of type gain, that is, x Ɒa y ?
f ~ x, a! ⱖ f ~ y, a! (according to increasing preference) or x Ɒa y ? f ~ x, a! ⱕ
f ~ y, a! (according to decreasing preference), where a 僆 AT, x, y 僆 U. For a
subset of attributes A 債 AT, we define x ⱰA y ? ∀ a 僆 A, x Ɒa y. That is, x
is at least as good as y with respect to all attributes in A. In general, the domain of
the condition criterion may be also discrete, but the preference order between its
values has to be provided.
The dominance relation that identifies granules of knowledge is defined as
follows.
For a given OIS, we say that x dominates y with respect to A 債 AT if
x ⱰA y, and denoted by xR Aⱖ y. Namely,
R Aⱖ ⫽ $~ y, x! 僆 U ⫻ U 6 y ⱰA x%
If ~ y, x! 僆 R Aⱖ , then y dominates x with respect to A.
Given A 債 AT and A ⫽ A 1 艛 A 2 , where attributes set A 1 according to
increasing preference, A 2 according to decreasing preference. The granules of
knowledge induced by the dominance relation R Aⱖ are the set of objects dominat-
ing x,
@x# ⱖ
A ⫽ $ y 僆 U 6 f ~ y, a 1 ! ⱖ f ~ x, a 1 !~∀ a 1 僆 A 1 !

and f ~ y, a 2 ! ⱕ f ~ x, a 2 !~∀ a 2 僆 A 2 !%
⫽ $ y 僆 U 6 ~ y, x! 僆 R Aⱖ %
and the set of objects dominated by x,
@x# ⱕ
A ⫽ $ y 僆 U 6 f ~ y, a 1 ! ⱕ f ~ x, a 1 !~∀ a 1 僆 A 1 !

and f ~ y, a 2 ! ⱖ f ~ x, a 2 !~∀ a 2 僆 A 2 !%
⫽ $ y 僆 U 6 ~ x, y! 僆 R Aⱖ %
which are called the A-dominating set and the A-dominated set with respect to
x 僆 U, respectively.
Let U/R Aⱖ denote classification, which is the family set $@x# ⱖ
A 6x 僆 U %. Any
element from U/R Aⱖ will be called a dominance class. Dominance classes in U/R Aⱖ
do not constitute a partition of U in general. They may overlap.
In the following, for simplicity, without any loss of generality, we only con-
sider condition attributes with increasing preference.
One can easily conclude the following property.

Property 1. Let R Aⱖ be a dominance relation.

(1) R Aⱖ is reflexive, transitive, and unsymmetric, so it is not a equivalence relation,

(2) if B 債 A 債 AT, then R Bⱖ 傶 R Aⱖ 傶 R AT ⱖ
,
(3) if B 債 A 債 AT, then @x# Bⱖ 傶 @x# Aⱖ 傶 @x# AT ⱖ
,
16 SHAO AND ZHANG
(4) if x j 僆 @x i # Aⱖ , then @x j # Aⱖ 債 @x i # Aⱖ and @x i # Aⱖ ⫽ 艛 $@x j # Aⱕ : x j 僆 @x i # Aⱖ %,
(5) @x i # Aⱖ ⫽ @x j # Aⱖ iff f ~ x i , a! ⫽ f ~ x j , a!~∀ a 僆 A!,
(6) J ⫽ $@x# Aⱖ 6x 僆 U % constitute a covering of U.

For any X 債 U and A 債 AT, the lower and upper approximations of X with
respect to the dominance relation R Aⱖ are defined as follows:
R Aⱖ ~X ! ⫽ $x 僆 U 6 @x# ⱖ ⱖ ⱖ
A 債 X %, R A ~X ! ⫽ $x 僆 U6@x# A 艚 X ⫽ ⭋%

Property 2. Let ~U, AT,V, f ! be an OIS and X,Y 債 U; its lower and upper
approximation satisfy the following properties:

(1) R Aⱖ ~X ! ⫽ ;R Aⱖ ~;X !, R Aⱖ ~X ! ⫽ ;R Aⱖ ~;X !

(2) R Aⱖ ~U ! ⫽ U, R Aⱖ ~⭋! ⫽ ⭋
(3) R Aⱖ ~X 艚 Y ! ⫽ R Aⱖ ~X ! 艚 R Aⱖ ~Y !, R Aⱖ ~X 艛 Y ! ⫽ R Aⱖ ~X ! 艛 R Aⱖ ~Y !
(4) R Aⱖ ~X 艛 Y ! 傶 R Aⱖ ~X ! 艛 R Aⱖ ~Y !, R Aⱖ ~X 艚 Y ! 債 R Aⱖ ~X ! 艚 R Aⱖ ~Y !
(5) R Aⱖ ~X ! 債 X 債 R Aⱖ ~X !
(6) R Aⱖ ~X ! 債 R Aⱖ ~R Aⱖ ~X !!, R Aⱖ ~R Aⱖ ~X !! 債 R Aⱖ ~X !

where ;X is the complement of X.

Example 1. An OIS is presented in Table I.

The dominance classes determined by AT are
ⱖ ⱖ
@x 1 # AT ⫽ $x 1 , x 2 , x 5 , x 6 %, @x 2 # AT ⫽ $x 2 , x 5 , x 6 %
ⱖ ⱖ
@x 3 # AT ⫽ $x 2 , x 3 , x 4 , x 5 , x 6 %, @x 4 # AT ⫽ $x 4 , x 6 %
ⱖ ⱖ
@x 5 # AT ⫽ $x 5 %, @x 6 # AT ⫽ $x 6 %
If X ⫽ $x 2 , x 3 , x 5 %, then
R Aⱖ ~X ! ⫽ $x 5 % 債 X, R Aⱖ ~X ! ⫽ $x 1 , x 2 , x 3 , x 5 % 傶 X

One can easily notice the property

R Aⱖ ~X ! ⫽ 艛 $@x# AT
ⱖ ⱖ
6 @x# AT 債 X %, R Aⱖ ~X ! ⫽ 艛 $@x# AT
ⱖ ⱖ
6 @x# AT 艚 X ⫽ ⭋%

does not hold.

Table I. An OIS.
U a1 a2 a3

x1 1 2 1
x2 3 2 2
x3 1 1 2
x4 2 1 3
x5 3 3 2
x6 3 2 3
 DOMINANCE RELATION AND RULES 17
It may happen that some of the attribute values for an object are missing.
To indicate such a situation a distinguished value, a so-called null value, is usu-
ally assigned to those attributes. We denote a special symbol * to indicate that
the value of an attribute is unknown. Here, we assume that an object x 僆 U
possesses only one value for an attribute a, a 僆 AT. Thus, if the value of an
attribute a is missing, then the real value must be one of the values of Va . In an
OIS, if ∀ a 僆 AT, ∀ x 僆 U, f ~ x, a! ⫽ *, then the OIS is called complete;
otherwise it is incomplete.

Example 2. An incomplete OIS is presented in Table II.

Let R A*ⱖ , A 債 AT, denote a binary dominance relation between objects that
are possibly dominant in terms of values of attributes set A. Let us define domi-
nance relation more precisely:

R A*ⱖ ⫽ $~ y, x! 僆 U ⫻ U 6 ∀ a 僆 A, f ~ y, a!
ⱖ f ~ x, a! or f ~ x, a! ⫽ * or f ~ y, a! ⫽ *%
By the definition of R A*ⱖ , it can be observed that if a pair of objects ~ y, x! from
U ⫻ U are in R A*ⱖ , then they are perceived as y dominates x; in other words, y
may have a better property than x with respect to A in reality.
Denoted by

A ⫽ $ y 僆 U 6 ~ y, x! 僆 R A %, @x# A ⫽ $ y 僆 U 6 ~ x, y! 僆 R A %
@x# *ⱖ *ⱖ *ⱕ *ⱖ

A describes objects that may dominate x and @x# A describes objects that may
@x# *ⱖ *ⱕ

be dominated by x in terms of A.
A 6 x 僆 U %.
Let U/R A*ⱖ denote classification, which is the family set $@x# *ⱖ
Any element from U/R A*ⱖ will be called a dominance class. All the dominance
classes in U/R A*ⱖ do not constitute a partition of U in general. They may overlap.
But, 艛 U/R A*ⱖ ⫽ U.

Table II. An incomplete OIS.

U a1 a2 a3

x1 3 2 2
x2 2 1 1
x3 2 1 *
x4 1 1 1
x5 1 * 1
x6 3 1 1
x7 2 1 1
x8 3 2 2
x9 2 1 2
x 10 1 1 2
x 11 2 1 2
x 12 3 * 2
18 SHAO AND ZHANG
Property 3. Let ~U, AT,V, f ! be an incomplete OIS and A 債 AT, then

R A*ⱖ ⫽ 艚R
a僆A
*ⱖ
$a%

Example 3. From Table II, we have

U/R AT
*ⱖ
⫽ $@x 1 # AT
*ⱖ
, @x 2 # AT
*ⱖ
, . . . , @x 12 # AT
*ⱖ
%
where

A ⫽ @x 8 # A ⫽ @x 12 # A ⫽ $x 1 , x 8 , x 12 %, @x 4 # A ⫽ @x 5 # A ⫽ U
@x 1 # *ⱖ *ⱖ *ⱖ *ⱖ *ⱖ

A ⫽ @x 3 # A ⫽ @x 7 # A ⫽ $x 1 , x 2 , x 3 , x 6 , x 7 , x 8 , x 9 , x 11 , x 12 %
@x 2 # *ⱖ *ⱖ *ⱖ

A ⫽ $x 1 , x 6 , x 8 , x 12 %, @x 10 # A ⫽ $x 1 , x 3 , x 8 , x 9 , x 10 , x 11 , x 12 %
@x 6 # *ⱖ *ⱖ

A ⫽ @x 11 # A ⫽ $x 1 , x 3 , x 8 , x 9 , x 11 , x 12 %
@x 9 # *ⱖ *ⱖ

Let I ⫽ ~U, AT,V, f ! be an incomplete OIS. For any X 債 U and A 債 AT,

the lower and upper approximations of X with respect to dominance relation R A*ⱖ
are defined as follows:
R A*ⱖ ~X ! ⫽ $x 僆 U 6 @x# *ⱖ
A 債 X % ⫽ $x 僆 X6 @x# A 債 X %
*ⱖ

A 艚 X ⫽ ⭋% ⫽ 艛 $@x# A 6 x 僆 X %
R A*ⱖ ~X ! ⫽ $x 僆 U 6 @x# *ⱖ *ⱖ

One can easily notice that R A*ⱖ ~X ! is a set of objects that belong to X with
certainty, whereas R A*ⱖ ~X ! is a set of objects that possibly belong to X.

Property 4. Let I ⫽ ~U, AT,V, f ! be an incomplete OIS, X 債 U, A 債 B 債

AT; then

(1) R A*ⱖ ~X ! 債 X 債 R A*ⱖ ~X !

(2) R A*ⱖ ~X ! 債 R B*ⱖ ~X !, RA*ⱖ ~X ! 傶 RB*ⱖ ~X !

3. ORDERED DECISION TABLES AND DECISION RULES

An ordered decision table (ODT) is an ordered information system I ⫽ ~U,
AT 艛 $d %,V, f !, where d ~d 僆 AT and * 僆 Vd ! is an overall preference called
the decision, and all the elements of AT are criterions. An ODT is called incom-
plete if it is an incomplete OIS; it is complete otherwise.
Furthermore, assume that the decision attribute d makes a partition of U
into a finite number of classes; let Cl ⫽ $Clt , t 僆 T %, T ⫽ $1, . . . , n%, be a set of
these classes that are ordered, that is, for all r, s 僆 T, such that r ⬎ s; the objects
from Clr are preferred to the objects from Cls . The sets to be approximated are an
upward union and a downward union of classes, which are defined respectively as
Cl tⱖ ⫽ 艛sⱖt Cls , Cl tⱕ ⫽ 艛sⱕt Cls , t ⫽ 1, . . . , n. The statement x 僆 Cl tⱖ means
“x belongs to at least class Clt ,” whereas x 僆 Cl tⱕ means “x belongs to at most
class Clt .”
 DOMINANCE RELATION AND RULES 19
The lower and upper approximations of Cl tⱖ with respect to dominance rela-
tion R A*ⱖ are defined as

R A*ⱖ ~Cl tⱖ ! ⫽ $x 僆 U 6 @x# *ⱖ ⱖ *ⱖ ⱖ

A 債 Cl t %, R A ~Cl t ! ⫽ 艛
x僆Cl tⱖ
A ,
@x# *ⱖ t ⫽ 1, . . . , n

Analogously, the lower and upper approximations of Cl tⱕ with respect to domi-

nance relation R A*ⱖ are defined as

R A*ⱖ ~Cl tⱕ ! ⫽ $x 僆 U 6 @x# *ⱕ ⱕ *ⱖ ⱕ

A 債 Cl t %, R A ~Cl t ! ⫽ 艛
x僆Cl tⱕ
A ,
@x# *ⱕ t ⫽ 1, . . . , n

The A-boundaries of Cl tⱖ and Cl tⱕ are defined as

BnA ~Cl tⱖ ! ⫽ R A*ⱖ ~Cl tⱖ ! ⫺ R A*ⱖ ~Cl tⱖ !,
BnA ~Cl tⱕ ! ⫽ R A*ⱖ ~Cl tⱕ ! ⫺ R A*ⱖ ~Cl tⱕ !, t ⫽ 1, . . . , n

In an OIS, an atomic expression (see Ref. 17) over a single attribute a is

defined as either ~a, ⱖ! (according to increasing preference) or ~a, ⱕ! (according
to decreasing preference). For any A 債 AT, an expression over A in OIS is defined
∧
by a僆A e~a!, where e~a! is an atomic expression over a. The set of all expres-
sions over A in an OIS is denoted by E~A!.

Example 4. In Example 1, AT ⫽ $a 1 , a 2 , a 3 %; consider the set of E~AT !:

E~$a 1 , a 2 , a 3 %! ⫽ $~a 1 , ⱖ! ∧ ~a 2 , ⱖ! ∧ ~a 3 , ⱖ!, ~a 1 , ⱖ! ∧ ~a 2 , ⱖ!
∧ ~a 3 , ⱕ!, . . . , ~a 1 , ⱕ! ∧ ~a 2 , ⱕ! ∧ ~a 3 , ⱕ!%
In an OIS, a 僆 AT, v1 僆 Va , an atomic formula over a single attribute a is
defined as either ~a,ⱖ, v1 ! (according to increasing preference) or ~a,ⱕ, v1 ! (accord-
ing to decreasing preference). For any A 債 AT, a formula over A in OIS is defined
∧
by a僆A m~a!, where m~a! is an atomic formula over a. The set of all formulas
over A in an OIS is denoted by M~A!. Let formula f 僆 M~A!, 7f7 denotes the set
of objects satisfying formula f. For example, ~a,ⱖ, v1 ! and ~a,ⱕ, v1 ! are atomic
formulas; then
7~a,ⱖ, v1 !7 ⫽ $x 僆 U 6 f ~ x, a! ⱖ v1 %, 7~a,ⱕ, v1 !7 ⫽ $x 僆 U 6 f ~ x, a! ⱕ v1 %

Definition 2. Consider a subset of attributes A 債 AT. For two formulas f 僆

M~A! and w 僆 M~d !, a decision rule, denoted by f r w, is read“if f then w.”
The formula f is called the rule’s antecedent, and the formula w is called the rule’s
consequent.

We say that an object supports a decision rule if it matches both the condition
and the decision parts of the rule. On the other hand, an object is covered by a
decision rule if it matches the condition parts of the rule.
A decision rule states how “evaluation of objects on attributes A is at least as
good as a given level” or “evaluation of objects on attributes A is at most as good
20 SHAO AND ZHANG
as a given level” determines “objects belong (or possibly belong) to at least a
given class” or “objects belong (or possibly belong) to at most a given class.”
There are four type of decision rules (see Refs. 14 and 24) that can be
considered:

(1) certain Dⱖ-decision rules with the following syntax:

if f ~ x, a 1 ! ⱖ va 1 and f ~ x, a 2 ! ⱖ va 2 and . . . f ~ x, a k ! ⱖ va k

and f ~ x, a k⫹1 ! ⱕ va k⫹1 and . . . f ~ x, a p ! ⱕ va p , then x 僆 Cl tⱖ

(2) possible Dⱖ-decision rules with the following syntax:

if f ~ x, a 1 ! ⱖ va 1 and f ~ x, a 2 ! ⱖ va 2 and . . . f ~ x, a k ! ⱖ va k

and f ~ x, a k⫹1 ! ⱕ va k⫹1 and . . . f ~ x, a p ! ⱕ va p , then x could belong to Cl tⱖ

(3) certain Dⱕ-decision rules with the following syntax:

if f ~ x, a 1 ! ⱕ va 1 and f ~ x, a 2 ! ⱕ va 2 and . . . f ~ x, a k ! ⱕ va k

and f ~ x, a k⫹1 ! ⱖ va k⫹1 and . . . f ~ x, a p ! ⱖ va p , then x 僆 Cl tⱕ

(4) possible Dⱕ-decision rules with the following syntax:

if f ~ x, a 1 ! ⱕ va 1 and f ~ x, a 2 ! ⱕ va 2 and . . . f ~ x, a k ! ⱕ va k

and f ~ x, a k⫹1 ! ⱖ va k⫹1 and . . . f ~ x, a p ! ⱖ va p , then x could belong to Cl tⱕ

where O1 ⫽ $a 1 , . . . a k % 債 AT, O2 ⫽ $a k⫹1 , . . . , a p % 債 AT, O ⫽ O1 艛 O2 , O1

with increasing preference and O2 with decreasing preference, ~va 1 , . . . , va p ! 僆
Va 1 ⫻ Va 2 ⫻ {{{ ⫻ Va k , t 僆 T.
For a given upward or downward union Cl tⱖ or Cl sⱕ , s, t 僆 T, the rules
induced under a hypothesis that objects belonging to R A*ⱖ ~Cl tⱖ ! or to R A*ⱕ ~Cl tⱕ !
are positive and all the others negative suggest the assignment of an object to “at
least class Clt ” or to “at most class Cls ,” respectively. The rules induced under a
hypothesis that objects belonging to R A*ⱖ ~Cl tⱖ ! or to R A*ⱕ ~Cl tⱕ ! are positive and
all the others negative suggest that an object could belongs to “at least class Clt ”
or to “at most class Cls ,” respectively.
Moreover, each decision rule should be minimal. Because a decision rule is
an implication, by a minimal decision rule we understand such an implication to
be that there is no other implication with an antecedent of at last the same weakness
(in other words, a rule using a subset of elementary condition or/and weaker ele-
mentary conditions) and a consequent of at least the same strength (in other words,
a rule assigning objects to the same union or subunion of class).

Example 5. Let us consider an incomplete ODT, constructed from an incomplete

OIS presented in Table II and extended by decision attribute d as shown in Table III.
 DOMINANCE RELATION AND RULES 21
Table III. An incomplete ODT.
U a1 a2 a3 d

x1 3 2 2 2
x2 2 1 1 1
x3 2 1 * 2
x4 1 1 1 1
x5 1 * 1 1
x6 3 1 1 2
x7 2 1 1 2
x8 3 2 2 2
x9 2 1 2 2
x 10 1 1 2 1
x 11 2 1 2 2
x 12 3 * 2 2

From the table we have Cl ⫽ $Cl1 , Cl2 %, where

Cl 1 ⫽ $x 2 , x 4 , x 5 , x 10 %, Cl 2 ⫽ $x 1 , x 3 , x 6 , x 7 , x 8 , x 9 , x 11 , x 12 %

Because only two classes are considered, we have Cl 1ⱕ ⫽ Cl1 and Cl 2ⱖ ⫽ Cl2 .
Thus

R AT
*ⱖ
~Cl 1ⱕ ! ⫽ $x 4 , x 5 , x 10 %, R AT
*ⱖ
~Cl 1ⱕ ! ⫽ $x 2 , x 3 , x 4 , x 5 , x 7 , x 10 %

R AT
*ⱖ
~Cl 2ⱖ ! ⫽ $x 1 , x 6 , x 8 , x 9 , x 11 , x 12 %,

R AT
*ⱖ
~Cl 2ⱖ ! ⫽ $x 1 , x 2 , x 3 , x 6 , x 7 , x 8 , x 9 , x 11 , x 12 %

BnAT
*
~Cl 1ⱕ ! ⫽ $x 2 , x 3 , x 7 %, BnAT
*
~Cl 2ⱖ ! ⫽ $x 2 , x 3 , x 7 %

We have the following set of minimal decision rules from the considered incom-
plete ordered decision table:

∧ ∧
r1 : ~a 1 ,ⱖ,3! ~a 2 ,ⱖ,1! ~a 3 ,ⱖ,1! r ~d,ⱖ,2! // supported by objects
x 1 , x 6 , x 8 , x 12
∧ ∧
r2 : ~a 1 ,ⱖ,3! ~a 2 ,⫽,*! ~a 3 ,ⱖ,1! r ~d,ⱖ,2! // supported by object x 12
∧ ∧
r3 : ~a 1 ,ⱖ,2! ~a 2 ,ⱖ,1! ~a 3 ,ⱖ,2! r ~d,ⱖ,2! // supported by objects
x 8 , x 9 , x 11
∧ ∧
r4 : ~a 1 ,ⱖ,2! ~a 2 ,⫽,*! ~a 3 ,ⱖ,2! r ~d,ⱖ,2! // supported by objects x 12
∧ ∧
r5 : ~a 1 ,ⱕ,1! ~a 2 ,ⱕ,1! ~a 3 ,ⱕ,2! r ~d,ⱕ,1! // supported by objects x 4 , x 10
∧ ∧
r6 : ~a 1 ,ⱕ,1! ~a 2 ,⫽,*! ~a 3 ,ⱕ,1! r ~d,ⱕ,1! // supported by objects x 5
∧ ∧
r7 : ~a 1 ,⫽,2! ~a 2 ,⫽,1! ~a 3 ,⫽,1! r ~d,ⱕ,1! ~d,ⱖ,2! // supported ∨
by objects x 2 , x 7
∧ ∧
r8 : ~a 1 ,⫽,2! ~a 2 ,⫽,1! ~a 3 ,⫽,*! r ~d,ⱕ,1! ~d,ⱖ,2! // supported ∨
by objects x 3
22 SHAO AND ZHANG
where r1 , r2 , r3 , r4 are certain Dⱖ-decision rules, r5 , r6 are certain Dⱕ-decision
rule, r7 , r8 are not only possible Dⱖ-decision rules but also possible Dⱕ-decision
rules.
With every ordering decision rule r : f r w, the certainty factor and cover-
age factor are defined respectively as

card~7f ∧ w7!
cerI ~f r w! ⫽
card~7f7!
card~7f ∧ w7!
covI ~f r w! ⫽
card~7w7!

The certainty factors can be interpreted as the frequency of objects having the
property w in the set of objects having the property f and the coverage factor as
the frequency of objects having the property f in the set of objects having the
property w.
For example, for the decision rule r1 :

~a 1 ,ⱖ,3! ∧ ~a ,ⱖ,1! ∧ ~a ,ⱖ,1! r ~d,ⱖ,2!

2 3

we have card~7f7! ⫽ 4, card~7f ∧ w7! ⫽ 4, card~7w7! ⫽ 8; therefore

cerI ~r1 ! ⫽ 1, covI ~r1 ! ⫽ 0.5.

4. KNOWLEDGE REDUCTION TO INCOMPLETE OIS AND ODT

In this section, the reducts of incomplete OIS and incomplete ODT are mate-
rialized by constructing a discernibility matrix.

4.1. The Reducts of Incomplete OIS

Let I ⫽ ~U, AT,V, f ! be an incomplete OIS and A 債 AT. We say that A is a

reduct of I if R A*ⱖ ⫽ R AT*ⱖ
and R B*ⱖ ⫽ R AT
*ⱖ
for any B 傺 A. Thus, a reduct of an
incomplete OIS is a minimal attribute subset satisfying R A*ⱖ ⫽ R AT *ⱖ
.
An attribute a 僆 AT is called dispensable with respect to R AT *ⱖ
if R AT
*ⱖ
⫽
R ~AT⫺$a%! ; a is called indispensable otherwise.
*ⱖ

Similar to the case of the equivalence relation, an incomplete OIS may have
many reducts. The set of all reducts of an incomplete OIS I is denoted by RED~AT !.
The set of all indispensable attributes is called the core with respect to domi-
nance relation R AT *ⱖ
and is denoted by CORE~AT !. An attribute in the core must be
in every reduct. So, CORE~AT ! ⫽ 艚 RED~AT !. The set of attributes CORE~AT !
are essential and cannot be eliminated. The core may be an empty set.
We denote

Dp ~ x, y! ⫽ $a 僆 A 6 ~ x, y! 僆 R $a%
*ⱖ
%
 DOMINANCE RELATION AND RULES 23
then Dp ~ x, y! is called the discernibility attribute set, and Dp ⫽ ~Dp ~ x, y! : x, y 僆
U ! is called the discernibility matrix of incomplete OIS. It is evident that
Dp ~ x, y! 艚 Dp ~ y, x! ⫽ ⭋ ~∀ x, y 僆 U !.

Property 5. Let I ⫽ ~U, AT,V, f ! be an incomplete OIS, A 債 AT, and Dp ~ x, y!

is the discernibility attributes set of I with respect to R AT
*ⱖ
; then A 艚 Dp ~ x, y! ⫽ ⭋
~Dp ~ x, y! ⫽ ⭋! iff R A ⫽ R AT .
*ⱖ *ⱖ

Proof. ~]!: For any ~ x, y! 僆 U ⫻ U, if ~ x, y! 僆 R AT *ⱖ

, then we have Dp ~ x, y! ⫽
⭋. Because A 艚 Dp ~ x, y! ⫽ ⭋, then there exists a 僆 A such that a 僆 Dp ~ x, y!,
that is, ~ x, y! 僆 R a*ⱖ . Hence ~ x, y! 僆 R A*ⱖ . Therefore R A*ⱖ 債 R AT
*ⱖ
. On the other
hand, it is evident that R AT 債 R A . Thus R A ⫽ R AT .
*ⱖ *ⱖ *ⱖ *ⱖ

~[!: Suppose R A*ⱖ ⫽ R AT *ⱖ

A ⫽ @x# AT ~∀ x 僆 U !. If y 僆 @x# AT ,
; then @x# *ⱖ *ⱖ *ⱖ

then y 僆 @x# A . Thus there exists a 僆 A such that ~ x, y! 僆 R $a% , and we have
*ⱖ *ⱖ

a 僆 Dp ~ x i , x j !. Therefore A 艚 Dp ~ x, y! ⫽ ⭋ ~∀ Dp ~ x, y! ⫽ ⭋!. 䡲

∨ ∨
Let Dp ~ x, y! be a Boolean expression that is equal to 1 if Dp ~ x, y! ⫽ ⭋.
∨
Otherwise, Dp ~ x, y! is a disjunction of variables corresponding to attributes con-
tained in Dp ~ x, y!.
∧ ∨ ∧ ∨
Let 䉭 ⫽ ~x, y!僆U⫻U Dp ~ x, y!, 䉭 ~ x! ⫽ y僆U Dp ~ x, y!; 䉭 is called a
discernibility function for incomplete OIS, 䉭 ~ x! is called a discernibility function
for object x.
Discernibility functions are monotonic Boolean functions and their prime
implications determine reducts uniquely.

Example 6. Determine all reducts for incomplete OIS presented in Table II.
Solution: Table IV is a discernibility matrix of incomplete OIS, where values
of Dp ~ x, y! for any pair ~ x, y! of objects from U are placed. From the Table IV, we
have

Table IV. The discernibility matrix of incomplete OIS.

x/y x1 x2 x3 x4 x5 x6 x7 x8 x9 x 10 x 11 x 12

x1
x2 a1 a2 a3 a1 a1 a2 a3 a3 a3 a3 a1 a3
x3 a1 a2 a1 a1 a2 a1
x4 a1 a2 a3 a1 a1 a1 a1 a1 a2 a3 a1 a3 a3 a1 a3 a1 a3
x5 a1 a3 a1 a1 a1 a1 a1 a3 a1 a3 a3 a1 a3 a1 a3
x6 a2 a3 a2 a3 a3 a3 a3 a3
x7 a1 a2 a3 a1 a1 a2 a3 a3 a3 a3 a1 a3
x8
x9 a1 a1 a1 a1
x 10 a1 a2 a1 a1 a1 a1 a1 a2 a1 a1 a1
x 11 a1 a2 a1 a1 a2 a1
x 12
24 SHAO AND ZHANG
䉭 ⫽ a 1 ∧ a 3 ∧ ~a 1 ∨ a 3 ! ∧ ~a 1 ∨ a 2 ! ∧ ~a 2 ∨ a 3 ! ∧ ~a 1 ∨ a 2 ∨ a 3 !
⫽ a1 ∧ a3
䉭 ~ x 1 ! ⫽ 䉭 ~ x 8 ! ⫽ 䉭 ~ x 12 ! ⫽ 1
䉭 ~ x 2 ! ⫽ a 1 ∧ a 3 ∧ ~a 1 ∨ a 3 ! ∧ ~a 1 ∨ a 2 ∨ a 3 ! ⫽ a 1 ∧ a 3
䉭 ~ x 3 ! ⫽ a 1 ∧ ~a 1 ∨ a 2 ! ⫽ a 1
䉭 ~ x 4 ! ⫽ a 1 ∧ a 3 ∧ ~a 1 ∨ a 3 ! ∧ ~a 1 ∨ a 2 ∨ a 3 ! ⫽ a 1 ∧ a 3
䉭 ~ x 5 ! ⫽ a 1 ∧ a 3 ∧ ~a 1 ∨ a 3 ! ⫽ a 1 ∧ a 3
䉭 ~ x 6 ! ⫽ a 3 ∧ ~a 2 ∨ a 3 ! ⫽ a 3
䉭 ~ x 7 ! ⫽ a 1 ∧ a 3 ∧ ~a 1 ∨ a 3 ! ∧ ~a 1 ∨ a 2 ∨ a 3 ! ⫽ a 1 ∧ a 3
䉭~ x9 ! ⫽ a 1
䉭 ~ x 10 ! ⫽ a 1 ∧ ~a 1 ∨ a 2 ! ⫽ a 1
䉭 ~ x 11 ! ⫽ a 1 ∧ ~a 1 ∨ a 2 ! ⫽ a 1

Therefore, $a 1 , a 3 % is a unique reduct for incomplete OIS, $a 1 % is a relative reduct

for objects x 3 , x 9 , x 10 , and x 11 , and $a 3 % is a relative reduct for objects x 6 .

4.2. The Reducts of Incomplete ODT

Let I ⫽ ~U, AT 艛 $d %,V, f ! be an incomplete ODT, d is an overall prefer-
ence of objects. Denoted by

R dⱖ ⫽ $~ x, y! : f ~d, x! ⱖ f ~d, y!%

R dⱖ is a dominance relation of decision attribute d.

I is called consistent if R AT
*ⱖ
債 R dⱖ ; otherwise, it is inconsistent.
Let I ⫽ ~U, AT 艛 $d %,V, f ! be a consistent incomplete ODT and A 債 AT.
We say that AT is a reduct of I if R A*ⱖ 債 R dⱖ and R B*ⱖ 債 R AT *ⱖ
for any B 傺 A.
We denote D ⫽ $~ x, y! : f ~d, x! ⬍ f ~d, y!%, defined by
*

Dp* ~ x, y! ⫽ 再⭋
$a 僆 AT : ~ x, y! 僆 R $a%
*ⱖ
% ~ x, y! 僆 D *
~ x, y! 僆 D *

then Dp* ~ x, y! is called a discernibility set for objects x and y, and Dp* ⫽
~Dp* ~ x, y! : x, y 僆 U ! is called a discernibility matrix for ODT I.
Similar to the incomplete OIS, we have the following property:

Property 6. Let I ⫽ ~U, AT 艛 $d %,V, f ! be a consistent incomplete ODT, A 債

AT, and Dp* ~ x, y! is the discernibility attributes set of I with respect to R dⱖ ; then
A 艚 Dp* ~ x, y! ⫽ ⭋ ~Dp* ~ x, y! ⫽ ⭋! iff R A*ⱖ 債 R dⱖ .

Proof. The proof is similar to the proof of Property 5. 䡲

*
∧
Let D ⫽ ~ x, y!僆U⫻U ∨ y!, D ~ x! ⫽ y僆U
Dp* ~ x, *
∧ ∨ Dp* ~ x,
y!; D is called a *

discernibility function for incomplete ODT; D* ~ x! is called a discernibility func-

tion for object x. D* determines reducts uniquely for the consistent incomplete
ODT; D* ~ x! determines relative reducts uniquely for object x.
 DOMINANCE RELATION AND RULES 25
Table V. The discernibility matrix of incomplete ODT.
x/y x1 x2 x3 x4 x5 x6 x7 x8 x9 x 10 x 11 x 12

x1
x2 a3
x3 a1 a2 a1 a1 a2 a1
x4 a1 a3
x5 a1 a3
x6 a2 a3 a1 a2 a3 a3 a3 a3
x7 a1 a2 a3 a1 a1 a2 a3 a3 a3 a3 a1 a3
x8
x9 a1 a2 a1 a1 a2 a1
x 10 a1
x 11 a1 a2 a3 a1 a1 a2 a1
x 12

Example 7. Determine all reducts for incomplete ODT presented in Table III by
computing prime implicants of discernibity functions D* .
Solution: Table V is a discernibility matrix of incomplete ODT, where values
of Dp* ~ x i , x j ! for any pair ~ x i , x j ! of objects from U are placed.
From Table V, we have

D* ⫽ a 1 ∧ a 3 ∧ ~a 1 ∨ a 2 ! ∧ ~a 1 ∨ a 3 ! ∧ ~a 2 ∨ a 3 ! ∧ ~a 1 ∨ a 2 ∨ a 3 !
⫽ a1 ∧ a3
䉭 * ~ x 1 ! ⫽ 䉭 * ~ x 8 ! ⫽ 䉭 * ~ x 12 ! ⫽ 1, 䉭 * ~ x 2 ! ⫽ a 3 , 䉭 * ~ x 10 ! ⫽ a 1
䉭 * ~ x 3 ! ⫽ 䉭 * ~ x 9 ! ⫽ a 1 ∧ ~a 1 ∨ a 2 ! ⫽ a 1 , 䉭 * ~ x 4 ! ⫽ 䉭 * ~ x 5 ! ⫽ a 1 ∧ a 3
䉭 * ~ x 6 ! ⫽ a 3 ∧ ~a 1 ∨ a 2 ! ∧ ~a 2 ∨ a 3 ! ⫽ a 1 ∧ a 3
䉭 * ~ x 7 ! ⫽ a 1 ∧ a 3 ∧ ~a 1 ∨ a 3 ! ∧ ~a 1 ∨ a 2 ∨ a 3 ! ⫽ a 1 ∧ a 3
䉭 * ~ x 11 ! ⫽ a 1 ∧ ~a 1 ∨ a 2 ! ∧ ~a 1 ∨ a 2 ∨ a 3 ! ⫽ a 1

Thus, $a 1 , a 3 % is a unique reduct for incomplete ODT and $a 1 % is a relative reduct

for objects x 3 , x 9 , x 10 , and x 11 .
The original minimal decision rules of incomplete ODT presented by Table III
can be reduced to the following set of minimal decision rules:

r1* : ~a 1 ,ⱖ,3! r ~d,ⱖ,2! // supported by objects x 1 , x 6 , x 8 , x 12

∧
r2* : ~a 1 ,ⱖ,2! ~a 3 ,ⱖ,2! r ~d,ⱖ,2! // supported by objects x 8 , x 9 , x 11
r3* : ~a 1 ,ⱕ,1! r ~d,ⱕ,1! // supported by objects x 4 , x 10
∧ ∨
r4* : ~a 1 ,⫽,2! ~a 3 ,⫽,1! r ~d,ⱕ,1! ~d,ⱖ,2! // supported by objects
x2 , x5 , x7
∧ ∨
r5* : ~a 1 ,⫽,2! ~a 3 ,⫽,*! r ~d,ⱕ,1! ~d,ⱖ,2! // supported by objects x 3

5. CONCLUSIONS
In this article, a method was proposed to handle incomplete ordered informa-
tion systems. The adaption concerns the new rough set approach based on the use
26 SHAO AND ZHANG
of the dominance relation. The proposed approach is an extension of the original
approach of the classical incomplete DT. They maintain the good characteristics
of the original one and boil down to the original one when dominance relation is
substituted by an equivalence relation. The proposed solution for rules extraction
and finding all reducts from an ordered incomplete information systems is simple.
This article provides a qualitative theoretical framework that may be important for
the analysis of ordered incomplete information systems.

Acknowledgments

The authors are grateful to the anonymous reviewers of this article for their
critical suggestions for improvements.

References

1. Pawlak Z. Rough sets. Int J Comput Inform Sci 1982;11:341—356.

2. Pawlak Z. Rough sets theory and its application to data analysis. Cybern Syst 1998;
29:661– 688.
3. An A, Shan N, Chan C, Cercone N, Ziarko W. Discovering rules for water demand predic-
tion: An enhanced rough-set approach. Eng Appl Artif Intell 1996;6:645– 653.
4. Slowinski R, Stefanowski J. Rough-set reasoning about uncertain date. Fund Inform
1996;27:229–244.
5. Skowron A. A synthesis of decision rules: Applications of discernibility matrix. In: Proc
Int Conf on Intelligent Information Systems, Augustow, Poland; 1993. pp 30– 46.
6. Shusaku T. Rule discovery in database with missing values based on rough set model. In:
Zhong N, Zhou L, editors. Methodologies for knowledge discovery and data mining. Bei-
jing, China: Springer; 1999. pp 274–278.
7. Pawlak Z. Rough sets: Theoretical aspects of reasoning about data. London: Kluwer Aca-
demic Publishers; 1991.
8. Polkowski L, Skowron A, editors. Rough sets in knowledge discovery 1,2. Methodology
and applications. Wurzburg: Physica-Verlag; 1998.
9. Greco S, Matarazzo B, Slowinski R. Rough approximation of a preference relation by
dominance relations. ICS Research Report 16/96, Warsaw: Warsaw University of Tech-
nology; 1996 and in Eur J Oper Res 1999;117:63–83.
10. Greco S, Matarazzo B, Slowinski R. A new rough set approach to multicriteria and multi-
attribute classification. In: Polkowski L, Skowron A, editors. Rough sets and current trends
in computing (RSTCTC ’98), Lecture Notes in Artificial Intelligence, vol 1424. Berlin:
Springer-Verlag; 1998. pp 60– 67.
11. Greco S, Matarazzo B, Slowinski R. A new rough set approach to evaluation of bank-
ruptcy risk. In: Zopounidis C, editor. Operational tools in the management of financial
risks. Dordrecht: Kluwer; 1999. pp 121–136.
12. Greco S, Matarazzo B, Slowinski R. Rough sets theory for multicriteria decision analysis.
Eur J Oper Res 2001;129:11– 47.
13. Greco S, Matarazzo B, Slowinski R. Rough sets methodology for sorting problems in
presence of multiple attributes and criteria. Eur J Oper Res 2002;138:247–259.
14. Greco S, Matarazzo B, Slowinski R, Stefanowski J. An algorithm for induction of decision
rules consistent with the dominance principle. In: Ziarko W, Yao Y, editors. Rough Sets
and Current Trends in Computing, RSCTC 2000, LNAI 2005; 2001. pp 304–313.
15. Dembczynski K, Pindur R, Susmaga R. Generation of exhaustive set of rules within
dominance-based rough set approach. Electron. Notes Theor Comput Sci 2003;82(4).
 DOMINANCE RELATION AND RULES 27
16. Dembczynski K, Pindur R, Susmaga R. Dominance-based rough set classifier without
induction of decision rules. Electron Notes Theor Comput Sci 2003;82(4).
17. Sai Y, Yao YY, Zhong N. Data analysis and mining in ordered information tables. Proc
2001 IEEE Int Conf on Data Mining. IEEE Computer Society Press; 2001. pp 497–504.
18. Kryszkiewicz M. Rough set approach to incomplete information systems. Inform Sci
1998;112:39– 49.
19. Kryszkiewicz M. Rules in incomplete information systems. Inform Sci 1999;113:271–292.
20. Roman S, Vanderpooten D. A generalized definition of rough approximations based on
similarity. IEEE Trans Knowl Data Eng 2000;12:331–336.
21. Akira N. A rough logic based on incomplete information and its application. Int J Approx
Reason 1996;15:367–378.
22. Greco S, Matarazzo B, Slowinski R. Handing missing values in rough set analysis of muti-
attribute and muti-criteria decision problems. In: Zhong N, Skowron A, Ohsuga S, editors.
New directions in rough sets, data mining and granular-soft computing. Lecture Notes in
Artificial Intelligence, vol. 1711. Berlin: Springer-Verlag; 1999. pp 146–157.
23. Skowron A. A synthesis of decision rules: Applications of discernibility matrix. In: Proc
Int Conf on Intelligence Information Systems, Augustow, Poland; 1993. pp 30– 40.
24. Slowinski R, Greco S, Matarazzo B. Rough set analysis of preference-ordered data. In:
Alpigini JJ, Peters JF, Skowronek J, Zhong N, editors. Rough Sets and Current Trends
in Computing, RSCTC 2002, LNAI 2475. Berlin, Heidelberg: Springer-Verlag; 2002.
pp 44–59.