Effective selective ensemble algorithms for support vector machines

Abstract

Listen

Translate article

Creating base learners with high diversity is an important principle in ensemble learning. Since support vector machines (SVM) tends to train consistent base classifiers by simply disturbing samples, it is a challenge to design effective ensemble algorithms for enhancing its generalization ability. In this paper, we present two novel selective ensemble algorithms for training SVM based on constraint projection technique and selective ensemble strategy. Diverse base classifiers are firstly generated in distinct projective spaces which determined upon random pairwise constraints, and then two selective ensemble approaches are applied to learn the optimal weights for aggregating them. Experimental results on UCI datasets indicate that the generalization ability of our proposed algorithms significantly outperform the existing ensemble ones, including Bagging, AdaBoost, feature Bagging and LoBag.