Applied Bayesian Modelling, 2nd Edition

2.5.2 Dirichlet process priors

Dirichlet Process Priors (abbreviated as DPP priors) offer another approach avoiding parametric assumptions and unlike a mixture of parametric densities, are less impeded by uncertainty about the appropriate number of sub-groups (Dey et al., 1999; Teh et al., 2006; Teh, 2010). The DPP method deals with possible clustering in the data without trying to specify the number of clusters, except perhaps a maximum conceivable number. Let be drawn from a distribution with unknown parameters , where a Dirichlet process prior is adopted for the , but a conventional parametric prior for the . The Dirichlet process involves a random distribution centred on a baseline prior from which candidate values for are drawn, and precision parameter describing the concentration ...