Variable Selection in the Kernel Machine Framework via the Garrote Kernel Selector

• Activity Number: 649
• Type: Invited
• Date/Time: Thursday, August 5, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Michael C. Wu*+
• Companies: The University of North Carolina at Chapel Hill
• Keywords: garrote kernel selector ; variable selection ; kernel machines ; high-dimensional data ; nonnegative garrote ; kernel ridge regression

Abstract

The complexity of high-throughput "omics" data requires the use of flexible methods that can account for non-linearity and complex interactions in building predictive models. Kernel machine methods, e.g. support vector machines and kernel ridge regression, meet these criteria and are frequently applied to build predictive models from genomic data. However, KM methods are still subject to considerable noise and decreased prediction accuracy when few predictors are related to the outcome. Variable selection is necessary. We propose the Garrote Kernel Selector (GKS), a statistical method for integrating regularization and variable selection with kernel machines that still maintains the flexible kernel formulation. Simulations and data applications show that the GKS can offer improvement over both sparse linear ridge regression and non-sparse kernel ridge regression.