The transcoding sampler for stick-breaking inferences on Dirichlet process mixtures (2304.02563v2)

Abstract: Dirichlet process mixture models suffer from slow mixing of the MCMC posterior chain produced by stick-breaking Gibbs samplers, as opposed to collapsed Gibbs samplers based on the Polya urn representation which have shorter integrated autocorrelation time (IAT). We study how cluster membership information is encoded under the two aforementioned samplers, and we introduce the transcoding algorithm to switch between encodings. We also develop the transcoding sampler, which consists of undertaking posterior partition inference with any high-efficiency sampler, such as collapsed Gibbs, and to subsequently transcode it to the stick-breaking representation via the transcoding algorithm, thereby allowing inference on all stick-breaking parameters of interest while retaining the shorter IAT of the high-efficiency sampler. The transcoding sampler is substantially simpler to implement than the slice sampler, it can inherit the shorter IAT of collapsed Gibbs samplers and it can also achieve zero IAT when paired with a posterior partition sampler that is i.i.d., such as the sequential importance sampler.