Improving bagging performance through multi-algorithm ensembles

Abstract

Bagging establishes a committee of classifiers first and then aggregates their outcomes through majority voting. Bagging has attracted considerable research interest and been applied in various application domains. Its advantages include an increased capability of handling small data sets, less sensitivity to noise or outliers, and a parallel structure for efficient implementations. However, it has been found to be less accurate than some other ensemble methods. In this paper, we propose an approach that improves bagging through the employment of multiple classification algorithms in ensembles. Our approach preserves the parallel structure of bagging and improves the accuracy of bagging. As a result, it unlocks the power and expands the user base of bagging. © 2012 Springer-Verlag.