Distance Dependent Maximum Margin Dirichlet Process Mixture

摘要

We propose distance dependent maximum margin Dirichlet Process Mixture (STANDPM), a nonparametric Bayesian clustering model that combines distance-based priors with the discriminatively learned likelihood of the Maximum Margin Dirichlet Process Mixture. STANDPM generalizes the distance-based prior introduced in the distance dependent Chinese Restaurant Process for non-sequential distances and allows modeling of complex dependencies between data points and clusters. The generalized distance-based prior is formulated as an abstract similarity measurement between a data point and a cluster. Empirical results show that the STANDPM model with abstract similarity achieves state-of-the-art performances on a number of challenging clustering datasets.