International Day Of Women in Statistics and Data Science

Generative data augmentation is widely used to address class imbalance by enriching minority classes with synthetic samples. Existing approaches typically employ a fixed augmentation strength across all classes, ignoring differences in class imbalance, data structure, and generative quality. We propose an adaptive augmentation framework that determines class-specific augmentation strengths using both class proportions and structural information measured through Gini correlation. This strategy allocates more synthetic data to underrepresented and weakly structured classes while limiting augmentation for well-represented classes.

We develop a theoretical framework showing that the resulting excess risk is controlled by a weighted combination of class-conditional Wasserstein discrepancies and Gini-based structural factors. We further establish consistency results and demonstrate that adaptive augmentation provides tighter control of risk distortion than fixed augmentation schemes.

Experiments on imbalanced classification datasets show consistent improvements in minority-class recall and macro-F1 performance, while empirical results closely match theoretical predictions. The proposed framework provides a principled, structure-aware foundation for generative data augmentation.