Enhancing K-means algorithm for solving classification problems

Abstract

K-means is the most popular clustering algorithm because of its efficiency and superior performance. However, the performance of K-means algorithm depends heavily on the selection of initial centroids. This paper proposes an extension to the original K-means algorithm enabling it to solve classification problems. First, the entropy concept is employed to adapt the traditional K-means algorithm to be used as a classification technique. Then, to improve the performance of K-means algorithm, a new scheme to select the initial cluster centers is proposed. The proposed models are tested on seven benchmark data sets from the UCI machine learning repository. Experimental results have shown that the proposed models outperform the learning vector quantization network in most of the tested data sets.