Experiments for the Number of Clusters in K-Means

Abstract

K-means is one of the most popular data mining and unsupervised learning algorithms that solve the well known clustering problem. The procedure follows a simple and easy way to classify a given data set through a pre-specified number of clusters K, therefore the problem of determining "the right number of clusters" has attracted considerable interest. However, to the authors' knowledge, no experimental results of their comparison have been reported so far. This paper presents results of such a comparison involving eight selection options presenting four approaches. We generate data according to a Gaussian-mixture distribution with clusters' spread and spatial sizes variant. Most consistent results are shown by the least squares and least modules version of an intelligent version of the method, iK-Means by Mirkin [14]. However, the right K is reproduced best by the Hartigan's [5] method. This leads us to propose an adjusted iK-Means method, which performs well in the current experiment setting.