Generalized Ewens–Pitman model for Bayesian clustering

Abstract

We propose a Bayesian method for clustering from discrete data structures that commonly arise in genetics and other applications. This method is equivariant with respect to relabelling units; unsampled units do not interfere with sampled data; and missing data do not hinder inference. Cluster inference using the posterior mode performs well on simulated and real datasets, and the posterior predictive distribution enables supervised learning based on a partial clustering of the sample.