Title: Augmented conditional sampler for nonparametric mixture models
Authors: Bernardo Nipoti - Trinity College Dublin (Ireland) [presenting]
Antonio Canale - University of Padua (Italy)
Riccardo Corradin - University of Milano Bicocca (Italy)
Abstract: Nonparametric mixture models based on the Pitman-Yor (PY) process are a flexible tool for density estimation and clustering. Two main classes of algorithms, namely marginal and conditionals, have been considered in literature. We propose a new algorithm, named Augmented Conditional Sampler (ACS), which, although technically conditional, is closely reminiscent of the Polya urn marginal scheme and features the same degree of interpretability. Unlike its most popular conditional competitors, the ACS does not rely on the stick-breaking representation of the underlying PY process and turns out to be more robust to the choice of the parameters characterising the distribution of the underlying PY process. The performance of the ACS is investigated and compared with popular competitors, by means of an extensive simulation study. Finally, the proposed sampler is used as the building block of a new algorithm for carrying out posterior inference based on a class of dependent nonparametric priors.