On Monte Carlo methods for the Dirichlet process mixture model, and the selection of its precision parameter prior

Abstract

Two issues commonly faced by users of Dirichlet process mixture models are: 1) how to appropriately select a hyperprior for its precision parameter alpha, and 2) the typically slow mixing of the MCMC chain produced by conditional Gibbs samplers based on its stick-breaking representation, as opposed to marginal collapsed Gibbs samplers based on the Polya urn, which have smaller integrated autocorrelation times.

In this thesis, we analyse the most common approaches to hyperprior selection for alpha, we identify their limitations, and we propose a new methodology to overcome them.

To address slow mixing, we revisit three label-switching Metropolis moves from the literature (Hastie et al., 2015; Papaspiliopoulos and Roberts, 2008), improve them, and introduce a fourth move. Secondly, we revisit two i.i.d. sequential importance samplers which operate in the collapsed space (Liu, 1996; S. N. MacEachern et al., 1999), and we develop a new sequential importance sampler for the stick-breaking parameters of Dirichlet process mixtures, which operates in the stick-breaking space and which has minimal integrated autocorrelation time. Thirdly, we introduce the i.i.d. transcoding algorithm which, conditional to a partition of the data, can infer back which specific stick in the stick-breaking construction each observation originated from. We use it as a building block to develop the transcoding sampler, which removes the need for label-switching Metropolis moves in the conditional stick-breaking sampler, as it uses the better performing marginal sampler (or any other sampler) to drive the MCMC chain, and augments its exchangeable partition posterior with conditional i.i.d. stick-breaking parameter inferences after the fact, thereby inheriting its shorter autocorrelation times.