EPIC: Efficient Integration of Partitional Clustering Algorithms for Classification

Abstract

Partitional algorithms form an extremely popular class of clustering algorithms. Primarily, these algorithms can be classified into two sub-categories: a) k-means based algorithms that presume the knowledge of a suitable k, and b) algorithms such as Leader, which take a distance threshold value, τ, as an input. In this work, we make the following contributions. We 1) propose a novel technique, EPIC, which is based on both the number of clusters, k and the distance threshold, τ, 2) demonstrate that the proposed algorithm achieves better performance than the standard k-means algorithm, and 3) present a generic scheme for integrating EPIC into different classification algorithms to reduce their training time complexity.