An empirical evaluation of common vector based classification methods and some extensions

Abstract

An empirical evaluation of linear and kernel common vector based approaches has been considered in this work. Both versions are extended by considering directions (attributes) that carry out very little information as if they were null. Experiments on different kinds of data confirm that using this as a regularization parameter leads to usually better (and never worse) results than the basic algorithms. © 2008 Springer Berlin Heidelberg.