Feature selection with a supervised similarity-based k-medoids clustering

Abstract

A supervised similarity-based k-medoids (SSKM) clustering algorithm is proposed for feature selection in classification problems. The set of original features is iteratively partitioned into k clusters, each of which is composed of similar features and represented by a feature yielding the maximum total of similarities with the other features in the duster. A supervised similarity measure is introduced to evaluate the similarity between two features for incorporating information of class labels of training patterns during clustering and representative selection. Experimental results show that our proposed method can select a more effective set of features for classification problems.