On the evolutionary design of heterogeneous Bagging models

Bagging is a popular ensemble algorithm based on the idea of data resampling. In this paper, aiming at increasing the incurred levels of ensemble diversity, we present an evolutionary approach for optimally designing Bagging models composed of heterogeneous components. To assess its potentials, experiments with well-known learning algorithms and classification datasets are discussed whereby the accuracy, generalization and diversity levels achieved with heterogeneous Bagging are matched against those delivered by standard Bagging with homogeneous components.