Synthetic Data Augmentation for Fairness under Data Imbalance: Theory and Adaptive Allocation
Authors: Zhengchi Ma, Kaifeng Zhang, Anru R. Zhang
Organizations: Department of Electrical & Computer Engineering, Duke University · Department of Statistical Science, Duke University · Department of Biostatistics & Bioinformatics and Department of Computer Science, Duke University
Data imbalance across sensitive groups and class labels is common in classification problems and can contribute to substantial fairness disparities. We study how such imbalance affects Equal Opportunity and how synthetic data augmentation can mitigate the resulting unfairness. We first identify two mechanisms, prevalence imbalance and group-size imbalance, through which changes in the training distribution shift the learned decision boundary and induce disparities in group true positive rates. Motivated by this analysis, we study exact full-balance augmentation, which equalizes all group-label cells. We establish conditions under which full balancing strictly improves Equal Opportunity together with a prediction-error comparison under the oracle balanced distribution, and show that the Equal Opportunity gap of the synthetic-data classifier converges to that of the oracle balanced classifier as the sample size grows and the generator error vanishes. The finite-sample convergence rate agrees, up to logarithmic factors and generator-approximation error, with the information-theoretic lower-bound rate under the corresponding regularity regimes. Since exact balancing need not provide the best fairness-accuracy trade-off in practice, we further propose Accuracy-Constrained Adaptive Synthetic Augmentation (ACASA), which learns both where and how much to augment by minimizing estimated fairness disparity subject to a constraint on predictive accuracy. We provide a finite-sample guarantee comparing the selected augmentation allocation with the population-optimal accuracy-feasible allocation. Simulation studies and real data application support the theoretical findings and demonstrate that synthetic augmentation can substantially reduce fairness disparities while achieving favorable empirical fairness-accuracy trade-offs.
Figures & tables
Figure 1: Overview of the main ideas and structure of the paper.
Figure 2: Illustration of how data imbalance can induce fairness disparities. Left: prevalence imbalance, with a scarce B+ cell. Right: group-size imbalance, with group B underrepresented. Dotted contours represent the underlying true distributions, and the solid line represents the learned decision boundary.
Figure 3: Comparison of the raw distribution with two balanced reference distributions.
Figure 4: Comparison of the raw distribution with the full-balance reference distribution.
Figure 5: Predictive performance under prevalence imbalance, measured by (a) accuracy, (b) F1 score, and (c) Matthews correlation coefficient (MCC). The horizontal axis shows the positive-class odds difference ρA−ρB . Curves show means, and shading denotes pointwise 95% bootstrap confidence intervals.
Figure 6: Predictive performance under group-size imbalance, measured by (a) accuracy, (b) F1 score, and (c) Matthews correlation coefficient (MCC). The horizontal axis shows α−1/2 . Curves show means, and shading denotes pointwise 95% bootstrap confidence intervals.
Figure 7
Figure 9: Equal Opportunity gap across accepted ACASA iterations. Solid curves show mean ACASA trajectories, while dashed and dotted lines show full-balancing and raw-training benchmarks. Shading denotes pointwise 95% bootstrap confidence intervals. Runs that stop early retain their final values in subsequent plotted iterations.
Figure 10: Fairness–performance trade-offs measured by (a) accuracy, (b) F1 score, and (c) Matthews correlation coefficient (MCC). For each method, we select separate thresholds for groups A and B to minimize prediction error on the calibration set subject to an Equal Opportunity gap at most ϵ , and vary ϵ to obtain the curves.
Figure 11: Equal Opportunity gap on the COMPAS data before post-processing. Both panels report the test-set EO gap at the common classification threshold of 0.5 . Solid curves show mean ACASA trajectories for reweighting, SMOTE, and diffusion model, while dashed lines show raw training and the corresponding full-balancing methods. Results are averaged over 100 runs, and error bars give pointwise 95% bootstrap confidence intervals based on 5,000 resamples. Runs that stop early retain their final accepted allocation.
Figure 12: Predictive performance on the COMPAS data after group-specific threshold post-processing. The left panel reports test-set accuracy under accuracy-constrained ACASA, and the right panel reports test-set MCC under MCC-constrained ACASA. Solid curves show the final ACASA allocations, while dashed curves show raw training and full balancing. Results are averaged over 100 runs, and error bars give pointwise 95% bootstrap confidence intervals based on 5,000 resamples. The value ϵ=1 corresponds to unconstrained threshold selection, not to the absence of post-processing.
Synthetic data augmentation is widely used to mitigate class imbalance, but its theoretical effects on score-based classification remain poorly understood. This paper develops a framework for characterizing when synthetic minority augmentation can improve threshold-integrated and threshold-optimized metrics, including AUROC, AUPRC, best-threshold balanced accuracy, and best-threshold \F1 score. We separate the effect of augmentation into two components: a change in effective class weighting and a discrepancy between the synthetic and true minority distributions. Under well-specified score models, the raw estimator already targets the likelihood-ratio ordering, which is population-optimal for the metrics considered. Consequently, augmentation cannot provide a fundamental population-level improvement beyond possible finite-sample variance reduction, and may introduce additional bias through synthetic distributional error. We further establish minimax lower bounds showing that the raw estimator already achieves the optimal metric-regret rate in the well-specified regime. Under misspecification, however, augmentation can play a qualitatively different role: by changing the effective class balance, it can alter the restricted-class projection and correct ranking errors induced by the raw imbalanced objective. We provide explicit improvement bounds quantifying the roles of approximation error, finite-sample estimation error, and synthetic distributional error. Simulation studies corroborate the theory, demonstrating limited gains under well-specification and nontrivial but nonmonotone improvements under misspecification.
Zhengchi Ma, Pengfei Lyu, Anru R. Zhang
Department of Electrical & Computer Engineering, Duke University · Department of Biostatistics & Bioinformatics, Duke University · Department of Biostatistics & Bioinformatics and Department of Computer Science, Duke University
We revisit the fairness notion of disparate impact for synthetic data generation (SDG), that assesses whether the utility of generated records is the same across sensitive groups. Our approach departs from existing work on fair SDG, that address the problem of correcting for undue biases in the observed distribution, hence redefining SDG as learning a distribution that is not that of the real data. By contrast, non-disparate impact is notably achieved when the synthetic and real distributions are the same. We expose reasons why SDG may fail to reach that solution and discuss why approximation and estimation errors occur and can be disparate across groups. We notably look into the expressive power of SDG methods relative to distribution complexity, sampling errors due to group proportions, and estimation errors induced by differential privacy mechanisms. We illustrate cases of disparate impact on both artificial and real-world data, focusing on SDG methods that rely on probabilistic graphical models. We also introduce a strategy of learning group-wise SDG models, that constitutes a simple baseline for analyzing and in some settings mitigating cases of disparate impact.
Paul Andrey, Michaël Perrot, Batiste Le Bars +1
Univ. Lille, Inria, CNRS, Centrale Lille, UMR 9189 - CRIStal, F-59000 Lille, France
Class imbalance complicates probabilistic classification because standard training objectives emphasize majority-class performance. Synthetic oversampling can reduce imbalance, but discrepancies between the synthetic and target minority distributions may bias the fitted classifier, especially because synthetic samples depend on the observed data. We propose a bias-correction procedure that estimates the generator-induced loss discrepancy from a held-out subset of majority observations and transfers this correction to the minority class under a uniform bias-transfer condition. We establish finite-sample bounds for bias transfer and for the excess balanced risk of the resulting empirical risk minimizer, and characterize a regime in which SMOTE induces non-negligible loss bias. The framework can also be implemented in imbalanced multi-task learning and propensity-score estimation, with details provided in the Supplementary Material. Real data analyses show that the correction is most useful when synthetic distortion is appreciable.
Pengfei Lyu, Zhengchi Ma, Linjun Zhang +1
Department of Biostatistics & Bioinformatics, Duke University · Department of Electrical & Computer Engineering, Duke University · Department of Statistics, Rutgers University +1