A new information theoretic clustering algorithm for robust unsupervised classification

Abstract

A new information-theoretic, unsupervised clustering algorithm is proposed. The algorithm relies on the relationship between scatter properties of given features and information notions such as the entropy. It is shown that cluster centers can be effectively determined by exploiting information content of data cloud around it. The algorithm has the advantage of maintaining overall data space without modifying its structure, such as subtracting a portion from the main cluster. © 2013 Elsevier B.V., All rights reserved.