A new model selection method for SVM

Abstract

In this paper, a new learning method is proposed to build Support Vector Machines (SVMs) Binary Decision Functions (BDF) of reduced complexity and efficient generalization. The aim is to build a fast and efficient SVM classifier. A criterion is defined to evaluate the Decision Function Quality (DFQ) which blendes recognition rate and complexity of a BDF. Vector Quantization (VQ) is used to simplify the training set. A model selection based on the selection of the simplification level, of a feature subset and of SVM hyperparameters is performed to optimize the DFQ. Search space for selecting the best model being huge, Tabu Search (TS) is used to find a good sub-optimal model on tractable times. Experimental results show the efficiency of the method. © Springer-Verlag Berlin Heidelberg 2006.