On Evidence Weighted Mixture Classification

Abstract

Calculation of the marginal likelihood or evidence is a problem central to model selection and model averaging in a Bayesian framework. Many sampling methods, especially (Reversible Jump) Markov chain Monte Carlo techniques, have been devised to avoid explicit calculation of the evidence, but they are limited to models with a common parameterisation. It is desirable to extend model averaging to models with disparate architectures and parameterisations. In this paper we present a straightforward general computational scheme for calculating the evidence, applicable to any model for which samples can be drawn from the posterior distribution of parameters conditioned on the data. The scheme is demonstrated on a simple feature subset selection example.