Automatic Cluster Number Selection using a Split and Merge K-Means Approach

The k-means method is a simple and fast clustering technique that exhibits the problem of specifying the optimal number of clusters preliminarily. We address the problem of cluster number selection by using a k-means approach that exploits local changes of internal validity indices to split or merge clusters. Our split and merge k-means issues criterion functions to select clusters to be split or merged and fitness assessments on cluster structure changes. Experiments on standard test data sets show that this approach selects an accurate number of clusters with reasonable runtime and accuracy.