Title: Bayesian mixtures of multiple scale distributions
Authors: Florence Forbes - INRIA (France) [presenting]
Alexis Arnaud - INRIA (France)
Benjamin Lemasson - Inserm (France)
Emmanuel Barbier - Inserm (France)
Russel Steele - McGill (Canada)
Abstract: Multiple scale distributions are multivariate distributions that exhibit a variety of shapes not necessarily elliptical while remaining analytical and tractable. We consider mixtures of such distributions for their ability to handle non standard, typically non-Gaussian clustering tasks. We propose a Bayesian formulation of the mixtures and a tractable inference procedure based on variational approximation. The interest of such a Bayesian formulation is illustrated on an important mixture model selection task, which is the issue of selecting automatically the number of components. We derive promising procedures that can be carried out in a single-run, in contrast to the more costly comparison of information criteria.