Professional Documents
Culture Documents
Marcin Szelg
Institute of Computing Science, Pozna University of Technology, 60-965 Pozna, Poland
12.5.2010
Outline
Introduction
CBR-DRSA Methodology Basic Notions and Denitions Decision Rules Certain and Possible Fuzzy Classication
Conclusions
Introduction
General Motivations
Case-based reasoning (CBR) is a natural way in which people solve problems. It is a process of solving new problems based on the solutions of similar problems from the past. When calculating similarity of two objects, main diculty consists in aggregation of dierent criteria/attributes; usually such aggregation is performed arbitrary, using weights or aggregation operators like sum, average or distance metrics. Therefore, there is a need for multi-criteria/multi-attribute modelling method that allows to include domain knowledge, can handle possible inconsistencies in data, and avoids any aggregation operators.
works with heterogenous attributes nominal, ordinal and cardinal (no need of discretization), enables to infer decision rule model from decision table (disaggregation-aggregation paradigm).
Advantages of decision rules: comprehensible form of knowledge representation, can represent any function (more general than utility functions or binary relations), resistant to irrelevant attributes, do not require aggregation operators, support backtracing, can explain past decisions and predict future decisions.
CBR-DRSA Methodology
10
CBR-DRSA
The rst step consists in creation of similarity tables, one for each reference object (). At this stage chosen marginal similarity functions are used to calculate marginal similarities w.r.t. chosen reference objects. Dierent marginal similarity functions can be used, depending on the domain of attribute , {1, . . . , }. The minimal requirement that each such function : [0, 1] must satisfy is that (, ) = 1, where is a reference object.
11
1 () ( )+1 1 ( () ( ))2 +1
12
Example for fuzzy IRIS problem part of the similarity table created for reference object no. 36 (5.1,3.4,1.5,0.2|1.0,0.5,0.4), with = ()1 ( )+1 , for {1, . . . , 4}:
13
We consider each decision class (fuzzy set) separately from the other classes. For each such class, we identify possible upward/downward -cuts in the following way: = { : ( ) }, = { : ( ) }. For an upward -cut it is required that > min , while for a downward -cut , it is required that < max , where denotes a set of membership values to observed in data.
14
The dominance relation between pairs of objects (, ) and (, ), w.r.t. set of condition attributes is dened as: (, ) (, ) : (, ) (, )
15
In the pair (, ), is considered to be a reference object, while is called a limit object, because it conditions the membership of + in (, ) and (, ).
16
Denitions of contextual -upper approximations: + ()( ) = { (, )}, ()( ) = { (, )}. Denitions of contextual -boundaries: ()( ) = ()( ) ()( ), ()( ) = ()( ) ()( ).
17
18
Decision rules are induced in order to identify similarity-based patterns in data. All pairs of objects (, ), where belongs to lower approximation of some or , are basis for induction of certain rules. All pairs of objects (, ), where belongs to upper approximation of some or , are basis for induction of possible rules.
19
Exemplary certain decision rule, for upward union 0.8 , generated for decision class setosa: If - (, 3.4) 0.59 - (, 1.5) 0.91 then belongs to class setosa to degree at least 0.8. Exemplary possible decision rule, for downward union 0.5 , generated for decision class versicolor : If - (, 6.5) 0.5 then could belong to class versicolor to degree at most 0.5.
20
There is a problem which decision rules resulting from lower/upper approximations of -cuts , should be used for classication. The minimal set of certain/possible rules is non-unique. The choice of such a set is arbitrary and non-trivial. On the other hand, (explicit) generation of a set of all certain/possible rules is computationally hard.
21
Proposed approach: Certain knowledge minimal set of certain rules (MCR), generated from lower approximations of and , independently for each decision class . Possible knowledge minimal set of possible rules (MPR), generated from upper approximations of and , independently for each decision class . Only minimal certain/possible rules are taken into account. Decision rules are generated by VC-DomLEM [1,2] algorithm.
22
Fuzzy classication of a new (test) object is performed independently for each decision class . For each such class two suggestions can be calculated: certain suggestion, resulting from application of MCR, possible suggestion, resulting from application of MPR. Each suggestion is obtained using a VC-DRSA classier [5].
23
References
1 2 3
J. Baszczyski, R. Sowiski, M. Szelg, Sequential Covering Rule Induction Algorithm for Variable Consistency Rough Set Approaches. Submitted to Information Sciences in 2009. J. Baszczyski, R. Sowiski, M. Szelg, VC-DomLEM: Rule induction algorithm for variable consistency rough set approaches. Research Report RA-07/09, Pozna University of Technology, 2009. S. Greco, B. Matarazzo, R. Sowiski, Granular Computing for Reasoning About Ordered Data: the Dominance-Based Rough Set Approach. Chapter 15 [in]: W. Pedrycz, A. Skowron, V. Kreinovich (eds.), Handbook of Granular Computing. John Wiley & Sons, Chichester, 2008, pp. 347373. S. Greco, B. Matarazzo, R. Sowiski, Case-based reasoning using gradual rules induced from dominance-based rough approximations. [In]: G. Wang, T. Li, J. W. Grzymaa-Busse, D. Miao, A. Skowron, Y. Yao (eds.), Rough Sets and Knowledge Technology (RSKT 2008). Lecture Notes in Articial Intelligence, vol. 5009, Springer-Verlag, Berlin, 2008, pp. 268275. J. Baszczyski, S. Greco, R. Sowiski, Multi-criteria classication A new scheme for application of dominance-based decision rules. European Journal of Operational Research, 181(3), 2007, pp. 10301044. S. Greco, B. Matarazzo, R. Sowiski, Dominance-based Rough Set Approach to Case-Based Reasoning. [In]: V. Torra, Y. Narukawa, A. Valls, J. Domingo-Ferrer (eds.), Modelling Decisions for Articial Intelligence. Lecture Notes in Articial Intelligence, vol. 3885, Springer-Verlag, Berlin Heidelberg, 2006, pp. 718.
5 6
25