Detecting the Number of Clusters during Expectation-Maximization Clustering Using Information Criterion

Abstract

This paper presents an algorithm to automatically determine the number of clusters in a given input data set, under a mixture of Gaussians assumption. Our algorithm extends the Expectation- Maximization clustering approach by starting with a single cluster assumption for the data, and recursively splitting one of the clusters in order to find a tighter fit. An Information Criterion parameter is used to make a selection between the current and previous model after each split. We build this approach upon prior work done on both the K-Means and Expectation-Maximization algorithms. We also present a novel idea for intelligent cluster splitting which minimizes convergence time and substantially improves accuracy.