On Constructing and Pruning SVM Ensembles

This paper proposes an effective method for constructing and pruning support vector machine ensembles for improved classification performance. Firstly we propose a novel method for constructing SVM ensembles. Traditionally an SVM ensemble is constructed by the data sampling method; In our method, however,each individual SVM classifier is trained by using the same original training set, but with different kernel parameters.Compared to traditional SVM ensemble methods, our method need not to tune the kernel parameters for each individual SVM, thus the training of the SVM ensemble can be simplified considerably. Furthermore, we also propose several efficient method for pruning the constructed SVM ensembles. The proposed pruning methods cannot only simplify the SVM ensemble, but also improve its performance. A set of experiments were conducted to prove the efficiency and affectivity of our proposed approaches.