Analyzing l1-loss and l2-loss support vector machines implemented in PERMON toolbox

Abstract

This paper deals with investigating l1-loss and l2-loss l2-regularized Support Vector Machines implemented in PermonSVM – a part of our PERMON toolbox. The loss functions quantify error between predicted and correct classifications of samples in cases of non-perfectly linearly separable classifications. In numerical experiments, we study properties of Hessians related to performance score of models and analyze convergence rate on 4 public available datasets. The Modified Proportioning and Reduced Gradient Projection algorithm is used as a solver for the dual Quadratic Programming problem resulting from Support Vector Machines formulations.