Online data clustering using variational learning of a hierarchical dirichlet process mixture of dirichlet distributions

Abstract

This paper proposes an online clustering approach based on both hierarchical Dirichlet processes and Dirichlet distributions. The deployment of hierarchical Dirichlet processes allows to resolve difficulties related to model selection thanks to its nonparametric nature that arises in the face of unknown number of mixture components. The consideration of the Dirichlet distribution is justified by its high flexibility for non-Gaussian data modeling as shown in several previous works. The resulting statistical model is learned using variational Bayes and is evaluated via a challenging application namely images clustering. The obtained results show the merits of the proposed statistical framework. © 2014 Springer-Verlag Berlin Heidelberg.