Model selection for the LS-SVM. Application to handwriting recognition

Support Vector Machine(SVM) is a powerful classifier used successfully in many pattern recognition problems. Furthermore, the good performance of SVM classifier has been shown in handwriting recognition field. Least Squares SVM, like SVM, is based on the marginmaximization principle performing structural risk, but its training is easier: it is only needed to solve a convex linear problem rather than the quadratic problem in SVM. In this paper, we propose to perform model selection for Least Squares SVM by using empirical error criterion. Experiments on handwriting character recognition show the usefulness of this classifier and demonstrate that LSSVM generalization performance is improved with model selection.