Prune the set of SV to improve the generalization performance of SVM

Abstract

Initiated by that the quality of training data may affect the model selection, this paper presents a method to improve the prediction performance of SVM through pruning the set of SV. That is, using a global comparable noise measure based on neighbor distribution information to identify noisy SVs, and weaken their role in training. The difference of this method from traditional one is that it need not to process noise for every instance in training set, and but only for those in SVs. The experiment result shows that when top noisy SVs are weakened the prediction performance of SVM is better for most categories.