Improving minority class prediction using cost-sensitive ensembles

In this paper, we address the problem of dealing with unbalanced datasets in the context of classification, i.e. where some of the classes contain significantly more objects than the other(s). We show that we can this problem by choosing classifiers for a committee of multiple classifier systems. In particular, we propose to design such an ensemble on the basis of a cost of elementary classifiers, given by a cost matrix. To assure the diversity of the ensemble each of the base classifiers is trained on a random subspace. This allows to improve the recognition rate of the minority class, which is typically low when using canonical classifiers. We evaluated our proposed algorithm on a variety of benchmark datasets and show that it significantly outperforms the base cost-sensitive classifier and its boosted version. The results confirm that our approach is a useful tool for dealing with unbalanced datasets.