Decomposition algorithms for training large-scale semiparametric support vector machines

Abstract

We describe a method for solving large-scale semiparametric support vector machines (SVMs) for regression problems. Most of the approaches proposed to date for large-scale SVMs cannot accommodate the multiple equality constraints that appear in semiparametric problems. Our approach uses a decomposition framework, with a primal-dual algorithm to find an approximate saddle point for the min-max formulation of each subproblem. We compare our method with algorithms previously proposed for semiparametric SVMs, and show that it scales well as the number of training examples grows. © 2009 Springer Berlin Heidelberg.