Modified NSGA-II Based Fuzzy Clustering of Categorical Attributes

The problem of clustering categorical data, whereno natural ordering among the elements of a categoricalattribute domain can be found, has been recently gainingsignificant attention from researchers. However, most of thesemethods attempt to optimize a single measure of the clusteringgoodness. Often, such a single measure may not be appropriatefor different kinds of datasets. In this paper a probabilitydensity multi-objective genetic algorithm-based approach forfuzzy clustering of categorical data is proposed that encodesthe cluster modes and simultaneously optimizes fuzzycompactness and fuzzy separation of the clusters. Here we usepopulation based incremental learning algorithm (PBIL) thatcan be considered as one of the simplest estimation ofdistribution algorithms (EDAs) in NSGA-II. Hence, wecompletely abandon the traditional crossover and mutationoperators of NSGA-II and reproduce new candidateindividuals through sampling from an estimated density ofpromising individuals in the current population and we calledthis method PNSGA-II. A statistical test of significance hasbeen conducted to establish the superiority of the proposedmulti-objective approach

Keywords:

component , PNSGA-II , fuzzy clustering , multi-objective genetic algorithm , Pareto optimality , NSGA-II , PBIL , EDAs