Variable Selection in Finite Mixture Models

• Activity Number: 416
• Type: Contributed
• Date/Time: Wednesday, August 5, 2009 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Volodymyr Melnykov*+ and Ranjan Maitra
• Companies: Iowa State University and Iowa State University
• Keywords: variable selection ; finite mixture models

Abstract

This paper provides methodology for identifying the important variables in Gaussian finite mixture models. Strategies for forward, backward as well as stepwise selection are detailed. In the first case, we start with one variable and sequentially test for significant improvement of the fitted clustering model, adding one component at a time. In the backward selection approach, we start with the full model and keep on dropping variables until the reduced model results in a significant drop in log-likelihood. The stepwise selection procedure combines these two approaches. The elimination of unimportant variables is seen to dramatically improve performance of model-based clustering algorithms on both simulation and standard classification data sets.