Unsupervised Ensemble Learning for Class Imbalance Problems

Ensemble learning, which aggregates multiple base (weak) learners to obtain a strong learner, is an effective approach for improving the generalization performance of a machine learning model. Several completely unsupervised ensemble learning approaches have been proposed in the literature for binary classification. However, most of them only considered the case that the two classes are balanced, and hence their performances deteriorate when there is significant class imbalance, which often happens in practice. This paper proposes a spectral meta-learner for class imbalance (SMLCI) approach to explicitly consider the class imbalance. Experiments on 12 UCI datasets from various domains verified that SMLCI significantly outperformed the individual base classifiers, and also five existing unsupervised ensemble learning approaches, when the balanced classification accuracy is used as the performance measure.