Localized projection learning

Abstract

It is interesting to compare different criteria of kernel machines. In this paper, the following is made: 1) to cope with the scaling problem of projection learning, we propose a dynamic localized projection learning using k nearest neighbors, 2) the localized method is compared with SVM from some viewpoints, and 3) approximate nearest neighbors are demonstrated their usefulness in such a localization. As a result, it is shown that SVM is superior to projection learning in many classification problems in its optimal setting but the setting is not easy. © 2010 Springer-Verlag Berlin Heidelberg.