Combining metric learning with multi-ratio undersampling to achieve imbalanced classification

Abstract

One problem that impairs the classification performance of several classification techniques is class
imbalance. One often used method to address the class imbalance is resampling; however, it still has a limited
data space, which further impairs performance. This research proposes MMEnsemble, an undersampling-based
unbalanced classification framework that integrates metric learning into a multi-ratio undersampling-based
ensemble, as a solution to this problem. Determining the proper sample ratio in the multi-ratio ensemble
approach is another issue that this system resolves. Twelve real-world datasets were used to assess it. In terms of
recall and ROC-AUC, it surpassed the state-of-the-art methods of metric learning, undersampling, and
oversampling; in terms of Gmean and F-measure metrics, it performed similarly to them.