Reference Prior Distributions in Bayesian Clustering Problems

• Activity Number: 551
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Russell Steele*+
• Companies: McGill University
• Keywords: Model-based clustering ; Bayesian inference ; Markov Chain Monte Carlo

Abstract

Bayesian model-based clustering methods are quite popular due to the ease with which Markov Chain Monte Carlo methods can be implemented for estimation and the ability to include prior information on the shape and size of the clusters. However, unlike in regular statistical models, the choice of prior distribution will always have an impact on inference for clustering models, particularly on inference for the selection of the number of components. In this talk, I will discuss the various possibilities for data-dependent reference priors for both the mixture component parameters and the mixing parameters of the mixture model. In particular, I will focus on the use of a generalized unit information prior for the mixture component parameters and a prior distribution on the mixing parameters that depends only on the number of observations.