Abstract
Class imbalance is a common and critical challenge in machine learning classification problems, resulting in low prediction accuracy. While numerous methods, especially data augmentation methods, have been proposed to address this issue, a method that works well on one dataset may perform poorly on another. To the best of our knowledge, there is still no one single best approach for handling class imbalance that can be uniformly applied. In this paper, we propose an approach named smart data augmentation (SDA), which aims to augment imbalanced data in an optimal way to maximize downstream classification accuracy. The key novelty of SDA is an equation that can bring about an augmentation method that provides a unified representation of existing sampling methods for handling multi-level class imbalance and allows easy fine-tuning. This framework allows SDA to be seen as a generalization of traditional methods, which in turn can be viewed as specific cases of SDA. Empirical results on a wide range of datasets demonstrate that SDA could significantly improve the performance of the most popular classifiers such as random forest, multi-layer perceptron, and histogram-based gradient boosting.