Particle Gibbs Split-Merge Sampling for Bayesian Inference in Mixture Models

Alexandre Bouchard-Côté; Arnaud Doucet; Andrew Roth

This paper presents an original Markov chain Monte Carlo method to sample from the posterior distribution of conjugate mixture models. This algorithm relies on a flexible split-merge procedure built using the particle Gibbs sampler introduced in Andrieu et al. (2009, 2010). The resulting so-called Particle Gibbs Split-Merge sampler does not require the computation of a complex acceptance ratio and can be implemented using existing sequential Monte Carlo libraries. We investigate its performance experimentally on synthetic problems as well as on geolocation data. Our results show that for a given computational budget, the Particle Gibbs Split-Merge sampler empirically outperforms existing split merge methods. The code and instructions allowing to reproduce the experiments is available at github.com/aroth85/pgsm.

Particle Gibbs Split-Merge Sampling for Bayesian Inference in Mixture Models

Abstract