Generating Diverse Ensembles to Counter the Problem of Class Imbalance

One of the more challenging problems faced by the data mining community is that of imbalanced datasets. In imbalanced datasets one class (sometimes severely) outnumbers the other class, causing correct, and useful predictions to be difficult to achieve. In order to combat this, many techniques have been proposed, especially centered around sampling methods. In this paper we propose an ensemble framework that combines random subspaces with sampling to overcome the class imbalance problem. We then experimentally verify this technique on a wide variety of datasets. We conclude by analyzing the performance of the ensembles, and showing that, overall, our technique provides a significant improvement.