Classifier ensembles for imbalanced classification

Abstract

Pattern classification algorithms learn from training samples to generalise to previously unseen data. In practice, many pattern classification tasks are imbalanced. That is, that there exist (many) more training samples of some classes compared to others. This in turn typically leads to poor classification performance on the minority classes even though often they are the more important ones. In this paper, we consider strategies for addressing class imbalance in the context of ensemble classifiers, i.e. classifiers that employ multiple predictors. In particular, we review approaches based on oversampling, an ensemble whose base classifiers are trained on balanced data subsets, a multiple classifier systems with cost-sensitive base classifiers, and a combination of one-class classifiers.