When Does Synthetic Data Augmentation Help in Imbalanced Learning?

Synthetic data are increasingly used to address data scarcity and imbalance in modern learning problems, especially when important outcomes are rare. In classification tasks with severe class imbalance, a common strategy is to generate synthetic minority samples and use them to augment the training set. Although effective in practice, the statistical benefits of this approach remain poorly understood: synthetic data augmentation may reduce the effects of imbalance, but it may also introduce bias if the generated samples do not accurately reflect the target distribution.

In this talk, I will present a statistical answer to the question in the title. Specifically, I will describe a unified theoretical framework that clarifies when synthetic data improve classification, when they may instead be harmful, and how the optimal amount of augmentation depends on generator quality and the underlying learning task. Our results show that the impact of synthetic data is determined by the interaction among class imbalance, synthetic bias, and the underlying loss geometry. These insights also yield a practical, data-driven method for choosing the amount of synthetic augmentation. If time permits, I will also briefly discuss how to directly correct synthetic bias, how synthetic data can improve fairness and threshold-independent metrics such as AUROC and AUPRC, and applications to echocardiogram diagnosis.