Ensemble Implementations on Diversified Support Vector Machines

Abstract

Support vector machine (SVM) is an effective algorithm in pattern recognition. But usually, standard SVM requires solving a quadratic program (QP) problem. In majority situations, most implementations of SVM are approximate solution to the QP problem. As the approximate solutions cannot achieve the expected performance of SRM theory, it is necessary to research ensemble methods for SVM. Recently, in order to augment the diversities of individual classifiers of SVM, many researchers use random partition with the whole training to form sub-training sets. Therefore the performance of aggregated SVM, which was trained on those subsets, was improved. We proposed the ensemble method based on different implementations of SVM, because they have large diversities by their different implementing methods. The experiment results showed that this method is effectively to improve the aggregated learner's performance.