Proximal SVM Ensemble Based on Feature Selection

Abstract

Ensemble is a very popular learning method. Among most of the existing approaches, Bagging is commonly used. However, Bagging is not very effective on the stable learners. Proximal SVM, a variant of SVM, is a stable learner, so Bagging does not work well for PSVM. For this, two new feature selection based PSVM ensemble methods are proposed, i.e. BRFS and BR. Through perturbing both the training set and the input features, component learners with high accuracy as well as high diversity can be obtained. The experimental results on four datasets from UCI demonstrate that the new approaches perform over a single PSVM and Bagging.