To SMOTE, or not to SMOTE?

In imbalanced binary classification problems the objective metric is often non-symmetric and associates a higher penalty with the minority samples. On the other hand, the loss function used for training is usually symmetric - equally penalizing majority and minority samples. Balancing schemes, that augment the data to be more balanced before training the model, were proposed to address this discrepancy and were shown to improve prediction performance empirically on tabular data. However, recent studies of consistent classifiers suggest that the metric discrepancy might not hinder prediction performance. In light of these recent theoretical results, we carefully revisit the empirical study of balancing tabular data. Our extensive experiments, on 73 datasets, show that generally, in accordance with theory, best prediction is achieved by using a strong consistent classifier and balancing is not beneficial. We further identity several scenarios for which balancing is effective and observe that prior studies mainly focus on these settings.