Parallel randomized support vector machine

Abstract

A parallel support vector machine based on randomized sampling technique is proposed in this paper. We modeled a new LP-type problem so that it works for general linear-nonseparable SVM training problems unlike the previous work [2]. A unique priority based sampling mechanism is used so that we can prove an average convergence rate that is so far the fastest bounded convergence rate to the best of our knowledge. The numerical results on synthesized data and a real geometric database show that our algorithm has good scalability. © Springer-Verlag Berlin Heidelberg 2006.