Bayesian Variable Selection for High-Dimensional Clustering

• Activity Number: 234
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Juanjuan Li*+
• Companies: Northwestern University
• Keywords: Bayesian variable selection ; Clustering analysis ; Gibbs sampler

Abstract

Variable selection and parameter reduction are two challenges for high-dimensional clustering with finite mixture models. Many of the previous variable selection procedures assume that the non-discriminating variables are independent of the discriminating variables. A more general variable selection model is proposed by assuming that some of the non-discriminating variables are linearly associated with the discriminating variables. A Bayesian algorithm is derived for this new model which can simultaneously conduct variable selection and clustering. For parameter reduction, the mean, covariance, and regression parameters of the model are integrated out and the marginal posterior function is used for the MCMC algorithm for the rest of the clustering parameters. The posterior inference and model identification are specified. Also the performance of the methodology is explored with simulated and real datasets.