Cancer Microarray Feature Selection Using Support Vector Machines: Comparing Regularization Techniques

• Activity Number: 562
• Type: Contributed
• Date/Time: Thursday, August 6, 2009 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Tim Peters*+ and David Bulger and Jean Y.H. Yang and To-ha Loi and David Ma
• Companies: Macquarie University and Macquarie University and University of Sydney and St. Vincent's Hospital and St. Vincent's Hospital
• Keywords: Microarray ; Machine learning ; Feature Selection ; Support Vector Machines ; Regularization ; Unbalanced datasets

Abstract

Microarray data set dimensionality reduction is a prerequisite for avoiding overfitting, and a precursor to developing diagnostic tools. We show that a generic stepwise regression using the linear support vector machine penalized margin width as the objective function, subject to regularization parameter grid-search, gives superior performance to three other feature-selection methods (least-angle regression, Random Forest, and stepwise regression on Fisher discriminants). We use a hierarchical validation method, applying leave-one-out cross-validation within the training subset, and applying the trained classifier to a separate test subset, on each of four two-class cancer-related gene expression data sets. The generic method shows superior results, especially on unbalanced data sets, including Golub (1999). A fixed regularization value appears nearly optimal for all four data sets