Bayesian Statistics

Download the full article

The Expectation-Maximisation applied to Variable Selection algorithm (EMVS) is a nice Bayesian selection tool. Its main feature is to lower considerably the computationnal power required to estimate posterior distributions in the feature space via MCMC. The EMVS is based on the continuous spike-and-slab normal mixture model and on a closed form expression of the EM algorithm. Variable selection is achieved through two important assumptions on the spike distribution, namely continuity and a positive variance parameter that introduces sparsity in the selection process. Once posterior modes have been discovered, model evaluation is carried out in a second step using a point mass spike distribution.