Feature Selection for Cluster Analysis: an Approach Based on the Simplified Silhouette Criterion

Abstract

This paper explores the problem of selecting relevant features for clustering, assuming that the number of clusters is not known a priori. The number of clusters and the subset of relevant features are usually inter-related. From this standpoint, we propose an exploratory data analysis method that considers the relationships between these two aspects. Empirical results in a number of synthetic and bioinformatics datasets show that the proposed approach can allow both reducing the number of features and providing good estimations of the number of clusters