Fuzzy clustering based on α-divergence for spherical data and for categorical multivariate data

This paper presents two clustering algorithms based on α-divergence between memberships and variables that control cluster sizes: one is for spherical data and the other for categorical multivariate data. First, this paper shows that a conventional method for vectorial data can be interpreted as the regularization of another conventional method with α-divergence. Second, with this interpretation, a spherical clustering algorithm based on α-divergence is derived from an optimization problem built by regularizing a conventional method with α-divergence. Third, this paper connects the facts that the α-divergence is a generalization of Kullback-Leibler (KL)-divergence, and that three conventional co-clustering methods are based on KL-divergence. Based on these facts, a co-clustering algorithm based on α-divergence is derived from an optimization problem built by extending the KL-divergence in conventional methods to α-divergence. This paper also demonstrates some numerical examples for the proposed methods.