Synthetic Augmentation in Imbalanced Learning: When It Helps, When It Hurts, and How Much to Add
Zhengchi Ma, Anru R. Zhang
TL;DR
This paper develops a unified, risk-based framework for synthetic minority augmentation in imbalanced learning, introducing the concept of a balanced population risk and decomposing the impact of augmentation into class-weight bias and generator-mismatch bias. It identifies two regimes—local asymmetry, where augmentation can help, and local symmetry, where it cannot and may hurt—along with three generator-quality settings (ideal, realistic, inconsistent). A central contribution is the explicit excess-risk representation that reveals a bias–variance trade-off as a function of synthetic size $\tilde{n}$ and generator accuracy, enabling principled tuning of $\tilde{n}$ via Validation-Tuned Synthetic Size (VTSS). Theoretical results are complemented by simulations and a sepsis-prediction case study showing when augmentation helps and how to tune its quantity, with VTSS often outperforming naive balancing. Practically, the work advocates treating synthetic size as a tunable hyperparameter and guides practitioners to adapt augmentation strategies to the direction and magnitude of generator-induced bias, improving minority-class performance without exacerbating bias from imperfect generators.
Abstract
Imbalanced classification, where one class is observed far less frequently than the other, often causes standard training procedures to prioritize the majority class and perform poorly on rare but important cases. A classic and widely used remedy is to augment the minority class with synthetic examples, but two basic questions remain under-resolved: when does synthetic augmentation actually help, and how many synthetic samples should be generated? We develop a unified statistical framework for synthetic augmentation in imbalanced learning, studying models trained on imbalanced data augmented with synthetic minority samples and evaluated under the balanced population risk. Our theory shows that synthetic data is not always beneficial. In a ``local symmetry" regime, imbalance is not the dominant source of error near the balanced optimum, so adding synthetic samples cannot improve learning rates and can even degrade performance by amplifying generator mismatch. When augmentation can help (a ``local asymmetry" regime), the optimal synthetic size depends on generator accuracy and on whether the generator's residual mismatch is directionally aligned with the intrinsic majority-minority shift. This structure can make the best synthetic size deviate from naive full balancing, sometimes by a small refinement and sometimes substantially when generator bias is systematic. Practically, we recommend Validation-Tuned Synthetic Size (VTSS): select the synthetic size by minimizing balanced validation loss over a range centered near the fully balanced baseline, while allowing meaningful departures when the data indicate them. Simulations and a real sepsis prediction study support the theory and illustrate when synthetic augmentation helps, when it cannot, and how to tune its quantity effectively.