Selective SVM Ensembles Based on Modified BPSO

Abstract

Selective ensemble is effective for improve the classification performance through taking full advantage of the diversity and supplement between base classifiers. A BPSO (binary particle swarm optimization) based selective SVM ensemble approach is proposed to ensure the diversity and supplement among base classifiers in the training phase and high performance in the selection phase. Firstly, bootstrap method introduced by Bagging is employed to select the training set; secondly, SVMs are trained with hyper-parameters randomly selected from the space defined with respect to the distribution characteristics of data sets; thirdly, taking classification accuracy of selected ensemble as the optimization object, BPSO is applied to acquire the final selective ensemble. Experiments indicate that the proposed approach remarkably improves the classification accuracy with much less member classifiers compare to the whole ensemble.