Comparison of methods for clustering of variables defined on the hypersphere.

Abstract

Methods of clustering of variables based on the identification of a mixture of Watson distributions defined on the hypersphere are considered. For the identification of this mixture, the following iterative methods: the Dynamic Clusters Method, the EM (Estimation-Maximisation) Algorithm and the Principal Cluster Component Analysis, are discussed and are compared using simulated and real data. The performance of the methods is compared for the same initial solution, by evaluating the final solutions obtained in these methods through the calculation of a between-groups variability measure and a within-groups variability measure.