Estimating the number of clusters using multivariate location test statistics

Abstract

In the cluster analysis, to determine the unknown number of clusters we use a criterion based on a classical location test statistic, Hotelling's T-2. At each clustering level, its theoretical threshold is studied in view of its statistical distribution and a multiple comparison problem. In order to examine its performance, extensive experiments are done with synthetic data generated from multivariate normal distributions and a set of real image data.