Professional Documents
Culture Documents
1-s2.0-S0377221701002442-main
1-s2.0-S0377221701002442-main
www.elsevier.com/locate/dsw
a
Faculty of Economics, University of Catania, Corso Italia, 55, 95129 Catania, Italy
b
Institute of Computing Science, Poznan University of Technology, 60-965 Poznan, Poland
Abstract
We consider a sorting (classification) problem in the presence of multiple attributes and criteria, called the MA&C
sorting problem. It consists in assignment of some actions to some pre-defined and preference-ordered decision classes.
The actions are described by a finite set of attributes and criteria. Both attributes and criteria take values from their
domains; however, the domains of attributes are not preference-ordered, while the domains of criteria (scales) are
totally ordered by preference relations. Among the attributes we distinguish between qualitative attributes and
quantitative attributes. In order to construct a comprehensive preference model that could be used to support the
sorting task, we consider preferential information of the decision maker (DM) in the form of assignment examples, i.e.
exemplary assignments of some reference actions to the decision classes. The preference model inferred from these
examples is a set of ‘‘if . . . , then . . . ’’ decision rules. The rules are derived from rough approximations of decision classes
made up of reference actions. They satisfy conditions of completeness and dominance, and manage with possible
ambiguity (inconsistencies) in the set of examples. Our idea of rough approximations involves three relations together:
indiscernibility, similarity and dominance defined on qualitative and quantitative attributes, and on criteria, respec-
tively. The usefulness of this approach is illustrated by an example. 2002 Published by Elsevier Science B.V.
Keywords: Rough sets; Sorting; Classification; Multiple criteria decision analysis; Decision rules
doubtful knowledge about the DM’s preferences is hesitation of the DM and unstable character of
done by distinction of different kinds of decision her/his preferences.
rules, depending on whether they are induced from 4. For a given set of decision examples, it is able to
lower approximations of decision classes or from determine what criteria and/or attributes are
the boundaries of these classes composed of in- relevant for approximation of the assignment
consistent examples. Such a preference model is made by the DM; such subsets of relevant crite-
more general than the classical functional or rela- ria and/or attributes are called reducts and their
tional model in MCDA and it is more under- intersection is called the core.
standable for the users because of its natural 5. It permits one to handle criteria and attributes
syntax (Greco et al., 2001b). A survey of the ex- jointly and, moreover, to consider two types
tended rough set approach to different MCDA of attributes: qualitative (symbolic, nominal,
problems has been recently presented in Greco like location) and quantitative (numerical, like
et al. (2001a). number of personnel).
In this paper, we wish to concentrate on the No other known approach to MCDA has all
extended rough set approach to one of major the above features. Our idea of extension of RST
classes of MCDA problems, which is the multiple- to deal with MA&C sorting problems consists in
criteria sorting problem. This problem is stated as building rough approximations of decision classes
follows: given a set of actions evaluated by a set of using a generalized ‘‘granule of knowledge’’. It is
criteria, assign these actions to some pre-defined defined by three relations considered jointly: in-
and preference-ordered decision classes, such that discernibility, defined on qualitative attributes,
each action is assigned to exactly one class. The similarity, defined on quantitative attributes, and
sorting problem is very similar to the problem of dominance, defined on criteria. The indiscernibility
classification; however, in the latter, the actions are relation is reflexive, symmetric and transitive,
evaluated by regular attributes and the decision while the similarity relation is reflexive and the
classes are not necessarily preference-ordered. The dominance relation is reflexive and transitive.
novelty of our present study is a joint consider- Decision rules with a special syntax are further
ation of criteria and regular attributes in one induced from these approximations of decision
sorting problem. Such a problem will be called an classes. The above idea builds on our previous
MA&C sorting problem. Although it has many study of MA&C sorting problems (Greco et al.,
counterparts in real life, it has not been considered 1998b), where only dominance and indiscernibility
either within MCDA or within methodologies relations were used together in the rough set ap-
dealing with multiple-attribute classification proach.
problems, like discriminant analysis, machine In Section 2, we define multicriteria and mul-
learning or neural networks. tiattribute rough approximations. Then, in Section
The interest of using the extended RST to 3, we characterize the syntax and semantics of
MA&C sorting problems is multiple indeed: decision rules induced from rough approxima-
1. It permits one to build the preference model of tions. Section 4 presents an illustrative example
a DM from a set of decision examples (s)he has solved by the original RST and by the extended
provided; these are exemplary assignments of RST. Section 5 contains conclusions.
some reference actions to the decision classes.
2. The preference model is in the form of a finite
set of specific decision rules, which ensures in- 2. Multicriteria and multiattribute rough approxi-
telligibility of the model; it also enables an easy mation
definition of possible strategies for passing from
one class to a better class through improvement For algorithmic reasons, knowledge about ref-
on some criteria. erence actions is represented in the form of an in-
3. When building the preference model, it does not formation table. The rows of the table are labelled
eliminate inconsistent examples that result from by reference actions, whereas columns are labelled
250 S. Greco et al. / European Journal of Operational Research 138 (2002) 247–259
by attributes and entries of the table are attribute relation, i.e. reflexive, symmetric and transitive;
values, called descriptors. Formally, by an infor- it concerns mainly qualitative attributes;
mation table we understand the 4-tuple 2. similarity relation Rq , which is only reflexive; it
S ¼ hU ; Q; V ; f i, where U is a finite set of reference
S concerns mainly quantitative attributes (see
actions, Q is a finite set of attributes, V ¼ q2Q Vq Slowinski and Vanderpooten, 1997, 2000).
and Vq is a domain of the attribute q, and f : U We introduce the following notation for disjoint
Q ! V is a total function such that f ðx; qÞ 2 Vq for subsets of C: C > is a subset of criteria, C ¼ is a
every q 2 Q, x 2 U , called an information function subset of qualitative attributes and C is a subset of
(see Pawlak, 1991). quantitative attributes. Therefore, C > [ C ¼ [
An information table can be seen as a decision C ¼ C and C > \ C ¼ ¼ ;, C > \ C ¼ ; and C ¼ \
table if in the set of attributes Q there are distin- C ¼ ;. Moreover, for any P C we denote by P > ,
guished two disjoint sets: set C of condition attri- P ¼ and P the following subsets of P , respectively:
butes and set D of decision attributes. • the subset of criteria for which outranking rela-
As was mentioned in Section 1, the notion of tions are considered, i.e. P > ¼ P \ C > ,
attribute differs from that of criterion because the • the subset of qualitative attributes for which in-
domain (scale) of a criterion has to be ordered discernibility relations are considered, i.e. P ¼ ¼
according to a decreasing or increasing preference, P \ C¼ ,
while the domain of the attribute does not have to • the subset of quantitative attributes for which
be ordered. Formally, for each q 2 C being a cri- similarity relations are considered, i.e. P ¼
terion there exists an outranking relation (Roy, P \C .
1985) Sq on the set of actions U such that xSq y For each x; y 2 U we say that action x domi-
means ‘‘x is at least as good as y with respect to nates action y on criteria from P > if and only if
criterion q’’. We suppose that Sq is a total pre-or- xSq y for all q 2 P > . The dominance relation is re-
der, i.e. a strongly complete and transitive binary flexive and transitive. Similarly, we say that x is
relation defined on U on the basis of evaluations indiscernible from y on attributes from P ¼ if and
f ð ; qÞ. If domain Vq of criterion q is quantitative only if xIq y for all q P ¼ , and x is similar to y on
and for each x; y 2 U , f ðx; qÞ P f ðy; qÞ implies attributes from P if and only if xRq y for all
xSq y, then Vq is a scale of preference of criterion q. qP .
If, however, for criterion q, Vq is not quantitative, Furthermore, assuming that the set of decision
or f ðx; qÞ P f ðy; qÞ does not imply xSq y, then in attributes D (possibly a singleton fdg) makes a
order to define a scale of preference of criterion q partition of the set of actions U into a finite number
one can choose a function gq : U ! R such that of decision classes, let Cl ¼ fClt ; t 2 T g,
for each x; y 2 U , xSq y if and only if gq ðxÞ P gq ðyÞ T ¼ f1; . . . ; ng, be a set of these classes such that
(see, e.g., Roubens and Vincke, 1985); passing each x 2 U belongs to one and only one Clt 2 Cl.
from f ðx; qÞ to gq ðxÞ for every x 2 U is called re- We suppose that the classes are ordered, i.e. for all
coding of the domain Vq . To recode Vq it is enough r; s 2 T , such that r > s, the actions from Clr are
to order the actions of U from the worst to the best preferred (strictly or weakly, Roy, 1985) to the ac-
on criterion q and to assign to gq ðxÞ consecutive tions from Cls . More formally, if S is a compre-
numbers corresponding to the rank of x in this hensive outranking relation on U , i.e. if for all
order, i.e. for z being the worst, gq ðzÞ ¼ 1, for w x; y 2 U , xSy means ‘‘x is at least as good as y’’, we
being the second worst, gq ðwÞ ¼ 2, and so on. suppose: ½x 2 Clr ; y 2 Cls ; r > s ) ½xSy and not
Then, the domain of function gq ð Þ becomes the ySx. The above assumptions are typical for con-
scale of preference of criterion q. sideration of a multiple-criteria sorting problem.
For each attribute q 2 C which is not a crite- In other words, the DM has assigned the ref-
rion, we admit a possibility of defining one of the erence actions to the classes of Cl according to the
two following binary relations on U : following comprehensive evaluation: the worst
1. indiscernibility relation Iq , which, according to actions are in Cl1 , the best actions are in Cln , the
the original rough sets theory, is an equivalence other actions belong to the remaining classes Clr ,
S. Greco et al. / European Journal of Operational Research 138 (2002) 247–259 251
according to the principle that the higher r 2 T , Given P C and x 2 U , the ‘‘granules of knowl-
the better class Clr . For example, in the risk as- edge’’ used for approximation in the extended
sessment problem, a financial manager can con- approach are:
sider Cl ¼ fCl1 ; Cl2 ; Cl3 g, where Cl1 is a class of • a set of actions y dominating action x on P > and
firms with ‘‘high risk’’, Cl2 is a class of firms with indiscernible from x on P ¼ , and such that x is
‘‘medium risk’’ and Cl3 is a class of firms with similar to y on P , DLþ P ðxÞ ¼ fy 2 U : yDP xg,
‘‘low risk’’. • a set of actions y dominating action x on P > and
The sets to be approximated are called the up- indiscernible from x on P ¼ , and such that y is
ward union and downward union of decision classes, similar to x on P , DUþ P ðxÞ ¼ fy 2 U : yDP xg,
respectively: • a set of actions y dominated by action x on P >
[ [ and indiscernible from x on P ¼ , and such that
CltP ¼ Cls ; Clt6 ¼ Cls ; t ¼ 1; . . . ; n:
sPt s6t
x is similar to y on P , DL
P ðxÞ ¼ fy 2 U : xDP yg,
• a set of actions y dominated by action x on P >
The statement x 2 CltP means ‘‘x belongs at least and indiscernible from x on P ¼ , and such that
to class Clt ’’, while x 2 Clt6 means ‘‘x belongs at y is similar to x on P ; DU
P ðxÞ ¼ fy 2 U : xDP yg.
most to class Clt ’’. Let us remark that Cl1P ¼ The sets DLþ ðxÞ Uþ
P and D P ðxÞ are called ‘‘in-sim-
Cln6 ¼ U , ClnP ¼ Cln and Cl16 ¼ Cl1 . Further- dominating’’ sets, and the sets DL U
P ðxÞ and DP ðxÞ,
more, for t ¼ 2; . . . ; n we have ‘‘in-sim-dominated’’ sets.
6
Clt1 ¼ U CltP and CltP ¼ U Clt1
6
: For any P C we say that x 2 U belongs to
CltP without any ambiguity if x 2 CltP and for each
The key idea of the rough set philosophy is action y 2 U dominating x on P > and indiscernible
approximation of one knowledge by another from x on P ¼ , and such that x is similar to y on P ,
knowledge. In the original approach, the knowl- also y belongs to CltP , i.e. DLþ P
P ðxÞ Clt . Fur-
edge being approximated is a partition of U into thermore, we say that y 2 U could belong to CltP if
decision classes generated by a set of decision at- there exists at least one action x 2 CltP such that y
tributes; the knowledge used for approximation is dominates x on P > , y is indiscernible from x on P ¼
a partition of U into elementary sets of actions and y is similar to x on P , i.e. y 2 DUþP ðxÞ.
that are indiscernible by a set of condition attri- Thus, with respect to P C, the set of all ac-
butes. The elementary sets are seen as ‘‘granules of tions belonging to CltP without any ambiguity
knowledge’’ used for approximation. constitutes the P -lower approximation of CltP ,
In our case, where among condition attributes denoted by P ðCltP Þ, and the set of all actions that
there are criteria, and decision classes are prefer- could belong to CltP constitutes the P -upper ap-
ence-ordered, the knowledge being approximated proximation of CltP , denoted by P ðCltP Þ:
is a collection of upward and downward unions of
P ðCltP Þ ¼ fx 2 U : DLþ P
P ðxÞ Clt g;
decision classes and the ‘‘granules of knowledge’’ [
are sets of actions defined using dominance, simi- P ðCltP Þ ¼ DUþ
P ðxÞ; for t ¼ 1; . . . ; n:
larity and indiscernibility relations together, in- x2CltP
stead of indiscernibility relation alone. This is the
main difference between the original rough set Analogously, using DL U
P ðxÞ and DP ðxÞ, one can
approach and the present extension. define P -lower approximation and P -upper ap-
For any P C let us define two reflexive binary proximation of Clt6 :
relations on U , denoted by DP and DP , and called P ðClt6 Þ ¼ fx 2 U : DL 6
P ðxÞ Clt g;
‘‘in-sim-dominance’’ relations, such that for each [
x; y 2 U : P ðClt6 Þ ¼ DU
P ðxÞ; for t ¼ 1; . . . ; n:
• xDP y if and only if xSq y for each q 2 P > , xIq y for x2Clt6
Due to monotonicity stated by Theorem 4, it is of qualitative attributes, and the last to similarity
possible to define the reduct and the core within on a subset of quantitative attributes.
the extended RST. Each minimal subset P C The following three types of decision rules can
such that cP ðClÞ ¼ cC ðClÞ is called a reduct of Cl be considered:
and denoted by REDCl ðCÞ. Let us remark that a • D P -decision rules which are generated from the
decision table can have more than one reduct. The lower approximation P ðCltP Þ, having the fol-
intersection of all reducts is called the core and lowing form:
denoted by CORECl ðCÞ. The algorithm for gen-
eration of all reducts in the dominance-based ‘‘if f ðx; q1 Þ P rq1 and . . . f ðx; qu Þ P rqu and
rough set approach has been presented in Susmaga f ðx; quþ1 Þ ¼ rquþ1 and . . . f ðx; qv Þ ¼ rqv and
et al. (1999). f ðx; qvþ1 Þ is similar to rqvþ1 and . . . f ðx; qz Þ is
similar to rqz , then x 2 CltP ’’, (STAT 1)
by the rules. Moreover, reassignment of an action One of three induction strategies can be adop-
from U matching some D P 6 -decision rules is ted to obtain a set of decision rules (Stefanowski
made to a set of contiguous classes being a union and Vanderpooten, 1994; Mienko et al., 1996;
of classes suggested by the rules. Stefanowski, 1998; Slowinski et al., 2000):
In consequence of using the asymmetric simi- • generation of a minimal description, i.e. a mini-
larity relation, the above conditions of completeness mal set of rules,
should be read as: the set of actions reassigned by • generation of an exhaustive description, i.e. all
the complete set of rules to original decision classes possible minimal rules for a given information
should be maximal and the assignment should be table,
the most precise possible. Let us explain what • generation of a characteristic description, i.e. a
‘‘maximal’’ and ‘‘most precise’’ means. This means set of minimal rules covering relatively many ac-
that no other rule can be induced which would re- tions each, however, altogether not necessarily
assign more actions to original classes or to smaller all actions from U .
sets of contiguous classes. For example, a set of rules
is not complete if it assigns action x to classes Cl2
and Cl3 , but there exists another rule, not included 4. An example
in this set, permitting precise assignment of x to Cl3 .
Analogously, a set of rules is not complete if it as- The following example (extending the one given
signs action x to classes Cl2 , Cl3 and Cl4 , but there by Pawlak, 1997) illustrates the concepts introduced
exists another rule, not included in this set, permit- above and shows the main differences between the
ting more precise assignment of x to Cl3 and Cl4 . original RST and the present extension. In Table 2,
The maximal set of actions that can be reclas- there are shown exemplary decisions of a DM con-
sified exactly by a complete set of rules can be cerning eight warehouses (reference actions) de-
formally characterized as follows: scribed by means of three condition attributes:
1. For each P C, the set LtP ðP Þ ¼ fw 2 U : • A1 , capacity of the sales staff,
DLþ P
and DUþ P • A2 , geographical region,
P ðwÞ Clt P ðwÞ Clt g is com-
posed of all prototypes of D P -decision rules in- • A3 , area.
volving criteria and attributes from P . The decision attribute d specifies the assignment
Therefore, one can reclassify by D P -decision made by the DM into two sets of warehouses
rules based on these prototypes all actions y making either profit or loss. Table 2 represents
such that there exists P C and w 2 LtP ðP Þ preferential information given by the DM.
for which yDP w.
2. For each P C, the set y 2 Lt6 ðP Þ ¼ 4.1. The results of the original rough set approach
fw 2 U : DL 6
P ðwÞ Clt and DP
U
ðwÞ Clt6 g
is composed of all prototypes of D 6 -decision Class Cl1 of the warehouses making loss and
rules involving criteria and attributes from P . class Cl2 of the warehouses making profit were
Therefore, one can reclassify by D 6 -decision
rules based on these prototypes all actions y
such that there exists P C and w 2 Lt6 ðP Þ Table 2
Decision table with exemplary decisions
for which wDP y.
All actions that cannot be reclassified exactly by Warehouse A1 A2 A3 d
D P - and D 6 -decision rules are assigned to sets of w1 Medium A 500 Loss
w2 Good A 400 Profit
contiguous classes by D P 6 -decision rules, so that
w3 Medium A 450 Profit
all actions are classified by a complete set of rules. w4 Good B 400 Loss
We call minimal each set of minimal decision w5 Good B 475 Profit
rules that is complete and non-redundant, i.e. ex- w6 Medium B 425 Profit
clusion of any rule from this set makes it non- w7 Medium B 350 Profit
w8 Medium B 350 Loss
complete.
S. Greco et al. / European Journal of Operational Research 138 (2002) 247–259 257
approximated by means of the original rough set C ¼ fA1 ; A2 ; A3 g and D ¼ fdg; however, we must
approach. It is clear that C ¼ fA1 ; A2 ; A3 g and observe that A1 is a criterion, while A2 and A3 are
D ¼ fdg. The C-lower approximations, the C-up- attributes. With respect to A1 , ‘‘good’’ is better
per approximations and the C-boundaries of sets than ‘‘medium’’, while with respect to A2 and A3 no
Cl1 and Cl2 are, respectively: preference is expressed on their domains. More-
over, for attribute A2 , the classical indiscernibility
CðCl1 Þ ¼ fw1; w4g; relation is considered, i.e. for each x; y 2 U , xIq y if
and only if f ðx; A2 Þ ¼ f ðy; A2 Þ, and for attribute
CðCl1 Þ ¼ fw1; w4; w7; w8g;
A3 , a similarity relation is defined such that for
BnC ðCl1 Þ ¼ fw7; w8g; each x; y 2 U , xRq y if and only if
decision rules. They satisfy conditions of com- Greco, S., Matarazzo, B., Slowinski, R., 2001a. Rough sets
pleteness, and dominance, and manage with pos- theory for multicriteria decision analysis. European Journal
of Operational Research 129, 1–47.
sible inconsistencies in the set of examples. Greco, S., Matarazzo, B., Slowinski, R., 2001b. Conjoint
Moreover, such a preference representation is measurement and rough set approach for multicriteria
more understandable for the users because of its sorting problems in presence of ordinal criteria. In: Colorni,
natural syntax. For this reason, it facilitates an A., Paruccini, M., Roy, B. (Eds.), AMCDA – Aide
interactive construction of the DM’s preference Multicritere a la Decision (Multiple Criteria Decision
Aiding). EUR Report, Joint Research Centre, The Euro-
model. pean Commission, Ispra, pp. 114–141.
Other extensions of RST for dealing with Grzymala-Busse, J.W, 1992. LERS – a system for learning from
problems of MCDA are described in Greco et al. examples based on rough sets. In: Slowinski, R. (Ed.),
(1999, 2000, 2001a). Intelligent Decision Support. Handbook of Applications
and Advances of the Rough Sets Theory. Kluwer Academic
Publishers, Dordrecht, pp. 3–18.
Mienko, R., Stefanowski J., Toumi, K., Vanderpooten, D.,
Acknowledgements 1996. Discovery-oriented induction of decision rules. Cahier
du LAMSADE no. 141, Universite de Paris Dauphine, Paris.
Pawlak, Z., 1991. Rough Sets. Theoretical Aspects of Reason-
The research of the first two authors has been ing about Data. Kluwer Academic Publishers, Dordrecht.
supported by the Italian Ministry of University Pawlak, Z., 1997. Rough set approach to knowledge-based
and Scientific Research (MURST). The third au- decision support. European Journal of Operational Re-
thor wishes to acknowledge financial support from search 99, 48–57.
Roubens, M., Vincke, Ph., 1985. Preference Modelling. Spring-
the State Committee for Scientific Research, grant
er, Berlin.
KBN 8T11F 006 19. Roy, B., 1985. Methodologie Multicritere d’Aide a la Decision.
Economica, Paris.
Slowinski, R. (Ed.), 1992. Intelligent Decision Support. Hand-
book of Applications and Advances of the Rough Sets
References Theory. Kluwer Academic Publishers, Dordrecht.
Slowinski, R., Stefanowski, J., Greco, S., Matarazzo, B., 2000.
Greco, S., Matarazzo, B., Slowinski, R., 1996. Rough approx- Rough sets based processing of inconsistent information in
imation of a preference relation by dominance relations. decision analysis. Control and Cybernetics 29, 379–404.
ICS Research Report 16/96, Warsaw University of Tech- Slowinski, R., Vanderpooten, D., 1997. Similarity relation as a
nology, Warsaw. European Journal of Operational Re- basis for rough approximations. In: Wang, P.P. (Ed.),
search 117, 63–83 (1999). Advances in Machine Intelligence and Soft Computing.
Greco, S., Matarazzo, B., Slowinski, R., 1998a. A new rough Duke University Press, Durham, NC, pp. 17–33.
set approach to evaluation of bankruptcy risk. In: Zo- Slowinski, R., Vanderpooten, D., 2000. A generalised definition
pounidis, C. (Ed.), Operational Tools in the Management of of rough approximations. IEEE Transactions on Data and
Financial Risks. Kluwer Academic Publishers, Dordrecht, Knowledge Engineering 12, 331–336.
pp. 121–136. Stefanowski, J., 1998. On rough set based approaches to
Greco, S., Matarazzo, B., Slowinski, R., 1998b. A new rough induction of decision rules. In: Polkowski, L., Skowron, A.
set approach to multicriteria and multiattribute classifica- (Eds.), Rough Sets in Knowledge Discovery, vol. 1. Physica-
tion. In: Polkowski, L., Skowron, A. (Eds.), Rough Sets and Verlag, Heidelberg, pp. 500–529.
Current Trends in Computing, Lecture Notes in Artificial Stefanowski, J., Vanderpooten, D., 1994. A general two stage
Intelligence, vol. 1424. Springer, Berlin, pp. 60–67. approach to rule induction from examples. In: Ziarko, W.
Greco, S., Matarazzo, B., Slowinski, R., 1999. The use of rough (Ed.), Rough Sets, Fuzzy Sets and Knowledge Discovery.
sets and fuzzy sets in MCDM. In: Gal, T., Stewart, T., Springer, Berlin, pp. 317–325.
Hanne, T. (Eds.), Advances in Multiple Criteria Decision Susmaga, R., Slowinski, R., Greco, S., Matarazzo, B., 1999.
Making. Kluwer Academic Publishers, Dordrecht, pp. 14.1– Computation of reducts for multi-attribute and multi-
14.59 (Chapter 14). criteria classification. In: Proceedings of the VIIth Work-
Greco, S., Matarazzo, B., Slowinski, R., 2000. Extension of the shop on Intelligent Information Systems. Institute of
rough set approach to multicriteria decision support. Computer Science Fundamentals, Polish Academy of Sci-
INFOR Journal (Canadian Journal of Operational Re- ences, Warsaw, pp. 154–163.
search and Information Processing) 38, 161–196.