Enriched Pitman-Yor processes
In Bayesian nonparametrics there exists a rich variety of discrete priors, including the Dirichlet process and its generalizations, which are nowadays well-established tools. Despite the remarkable advances, few proposals are tailored for modeling observations lying on product spaces, such as $\mathbb{R}^p$. Indeed, for multivariate random measures, most available priors lack flexibility and do not allow for separate partition structures among the spaces. We introduce a discrete nonparametric prior, termed enriched Pitman-Yor process (EPY), aimed at addressing these issues. Theoretical properties of this novel prior are extensively investigated. We discuss its formal link with the enriched Dirichlet process and normalized random measures, we describe a square-breaking representation and we obtain closed-form expressions for the posterior law and the involved urn schemes. In second place, we show that several existing approaches, including Dirichlet processes with a spike and slab base measure and mixture of mixtures models, implicitly rely on special cases of the EPY, which therefore constitutes a unified probabilistic framework for many Bayesian nonparametric priors. Interestingly, our unifying formulation will allow us to naturally extend these models while preserving their analytical tractability. As an illustration, we employ the EPY for a species sampling problem in ecology and for functional clustering in an e-commerce application.