A Novel Filter-Wrapper Algorithm on Intuitionistic Fuzzy Set for Attribute Reduction From Decision Tables

Abstract

Attribute reduction from decision tables is one of the crucial topics in data mining. This problem belongs to NP-hard and many approximation algorithms based on the filter or the filter-wrapper approaches have been designed to find the reducts. Intuitionistic fuzzy set (IFS) has been regarded as the effective tool to deal with such the problem by adding two degrees, namely the membership and non-membership for each data element. The separation of attributes in the view of two counterparts as in the IFS set would increase the quality of classification and reduce the reducts. From this motivation, this paper proposes a new filter-wrapper algorithm based on the IFS for attribute reduction from decision tables. The contributions include a new instituitionistics fuzzy distance between partitions accompanied with theoretical analysis. The filter-wrapper algorithm is designed based on that distance with the new stopping condition based on the concept of delta-equality. Experiments are conducted on the benchmark UCI machine learning repository datasets.

A New Approach for Attribute Reduction from Decision Table based on Intuitionistic Fuzzy Topology

“…FW IFT algorithm //Init phase; In steps 10-12: Calculate the Sig R F (c) base on formula (9), where:…”

“…In steps 16-24: Because ST = {{x 1 , x 2 }} ̸ = ∅, R F = ST.P OP = {x 1 , x 2 } and B R F ̸ = B I then go to steps 5-7 : Calculate the Sig R F (c) base on formula (9), where:…”

“…-The first is to compare the performance of the proposed algorithm with F IFPOS [10], F IFE [12], and F IFD [9] algorithms, where the F IFPOS algorithm uses the positive region (POS) measure based on the intuitionistic fuzzy rough set, the F IFE algorithm uses the intuitionistic fuzzy information granular entropy measure, and the F IFD algorithm uses the distance measure between the partitions based on the IFS framework.…”

“…There are two general feature selection strategies: wrappers [36] and filters [35]. While the wrapper strategy employs learning algorithms to evaluate selected attribute subsets, the Filter strategy selects attributes based on some measures such as information gain [24,26,27,28,29,33,34,39,40], consistency [1,23,25,30,47], distance [8,9,35,42,43], and dependency [27, 41,38,46]. These measures can be classified into distance and positive regions [30].…”

“…Results show that the size and accuracy of classification of the reduct are gained, but for the noisy datasets, the IFS were often opted due to the nature of some constraints in the IFS approximation space [3, ?]. Some IF measures were used, such as the IF positive region [10,11], IF entropy [12,13], and IF granular distance [9]. However, all methods mentioned above still use the measure to define the reduct, so that the size of the reduct is not optimal compared to the original datasets.…”