A novel selective ensemble algorithm based on approximate reducts and optimal sampling

Abstract

Ensemble techniques have been widely used for improving the classification accuracy, and recent studies show that ensembling classifiers through multi-modal perturbation can further improve the classification performance. In this paper, we propose a novel selective ensemble algorithm (called SE_AROS) based on approximate reducts and optimal sampling. In SE_AROS, we design a multi-modal perturbation method to generate different base classifiers. The proposed perturbation method can simultaneously perturb the attribute space and training set, which can increase the diversity of base classifiers. We compare SE_AROS with current ensemble algorithms of the same type on several UCI data sets, where the evidential KNN (k-nearest neighbors) classification algorithm is used to train base classifiers. Experimental results show that SE_AROS can provide competitive solutions for selective ensemble.