Aspects of Feature Selection in Functional Data

• Activity Number: 183
• Type: Invited
• Date/Time: Monday, August 7, 2006 : 2:00 PM to 3:50 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Philip J. Brown*+
• Companies: University of Kent
• Keywords: Bayesian methods ; hyper-LASSO ; wavelet functional modelling ; variable selection ; p>n problem

Abstract

We look at functional data as arising from infrared applications in chemometrics and mass spectroscopy data used in proteomics. The data may contain experimental factors and covariates, but there is a desire to discriminate between two or more groups. Modeling often is facilitated by the use of wavelets. We review a variety of approaches to modeling the functional data as response and modeling directly the discriminatory categories conditional on functional data and experimental factor/covariates. Our ultimate focus will be on Bayesian models that allow regularization. To this end, we look at a variety of forms of scale mixture of normal prior distributions, including forms of hyper-lasso and approaches to robustness and stability of discrimination. We are particularly interested in fast algorithms capable of scaling up to many variables and which are flexible.