DIMACS TR: 2007-08

Saddle Point Feature Selection In SVM Regression

Authors: Yuri Goncharov, Ilya Muchnik and Leonid Shvartser


SVM wrapper method, proposed in our previous work "Simultaneous Feature Selection and Margin Maximization Using Saddle Point Approach" is investigated and examined for the SVM regression. The method simultaneously maximizes margin and minimizes feature space with help of a modification of the standard criterion by adding to the basic objective function a third term, which directly penalizes a chosen set of variables. Our examination of the proposed min-max saddle point algorithm for the SVM regression case on a SAR Benchmark proves the ability of the introduced approach both to select small subspaces of features and to improve the regression prediction quality.

Paper Available at: ftp://dimacs.rutgers.edu/pub/dimacs/TechnicalReports/TechReports/2007/2007-08.ps.gz
DIMACS Home Page