Improving Fairness of AI Systems with Lossless De-biasing

Yan Zhou, Murat Kantarcioglu, Chris CliftonarXiv 2021 · arXiv

Introduction

As artificial intelligence systems are increasingly deployed in high-stakes domains such as criminal justice, lending, and healthcare, the issue of algorithmic bias has moved to the forefront of research. A persistent challenge in this field is the perceived "zero-sum game" between fairness and accuracy. Traditional de-biasing techniques often rely on "lossy" pre-processing—editing or removing sensitive data—which inevitably discards valuable information and degrades model performance.

This paper proposes an "information-lossless" approach to de-biasing. Rather than removing data to hide historical inequities, the authors suggest augmenting the training set with synthetic data to balance representation. By focusing on the scarcity of favorable outcomes for disadvantaged groups, this method aims to achieve fairness without the typical accuracy trade-offs.

Methodology: Information-Lossless Oversampling

The core of the proposed method is a strategic oversampling technique. Algorithmic bias often arises because certain demographic groups are underrepresented in the "favorable" outcome category (e.g., successful loan repayments or low recidivism risk). Instead of modifying existing records, the researchers use the Synthetic Minority Over-sampling Technique (SMOTE) to generate new, synthetic samples for the unprivileged group that reflect these favorable outcomes.

Oversampling the unprivileged favored group to balance the distribution.

This approach is theoretically grounded in the concept of $\mathcal{H}$-divergence. By adding synthetic samples, the researchers can mathematically bound the discrepancy between the distributions of privileged and unprivileged groups. Because the original data remains untouched and is only augmented, the process is considered "information-lossless," preserving the nuances required for high predictive accuracy.

Experimental Evaluation

The researchers evaluated their method against five standard fairness benchmarks: Adult (income prediction), COMPAS (recidivism), German Credit, Medical Expense, and Bank Marketing. The performance was compared against state-of-the-art de-biasing methods such as Reweighing and Prejudice Remover across various metrics, including Disparate Impact (DI) and Average Odds Difference.

Key Results

The results demonstrate that the lossless technique consistently matches or exceeds the performance of existing methods. On the Adult dataset, for instance, the technique significantly improved fairness metrics while maintaining a stable accuracy profile.

Performance of the synthesized classification approach, showing balanced accuracy and fairness metrics across thresholds.

As shown in the performance plots, the synthesized approach allows for a classification threshold that optimizes both balanced accuracy and fairness. Compared to the original biased model, the synthesized model reduces the bias metric (represented by the red line in the figures) significantly at the optimal accuracy point—effectively moving the bias score from approximately 0.6 down to 0.2. In many cases, the accuracy drop was less than 3%, and occasionally, the balanced data even improved generalization and overall performance.

Discussion and Conclusion

The "lossless" de-biasing approach shifts the perspective on algorithmic bias from a data-correction problem to a data-completion problem. By addressing the scarcity of favorable minority representation through synthetic augmentation, the method bypasses the traditional accuracy-fairness trade-off.

While highly effective, the authors note that the success of this technique depends on maintaining realistic class priors. Future research may explore using synthetic data to explicitly compensate for systemic biases that are not fully captured in raw historical data. Ultimately, this work provides a robust framework for building AI systems that are both highly accurate and socially equitable.

Summary updated

Abstract

In today's society, AI systems are increasingly used to make critical decisions such as credit scoring and patient triage. However, great convenience brought by AI systems comes with troubling prevalence of bias against underrepresented groups. Mitigating bias in AI systems to increase overall fairness has emerged as an important challenge. Existing studies on mitigating bias in AI systems focus on eliminating sensitive demographic information embedded in data. Given the temporal and contextual complexity of conceptualizing fairness, lossy treatment of demographic information may contribute to an unnecessary trade-off between accuracy and fairness, especially when demographic attributes and class labels are correlated. In this paper, we present an information-lossless de-biasing technique that targets the scarcity of data in the disadvantaged group. Unlike the existing work, we demonstrate, both theoretically and empirically, that oversampling underrepresented groups can not only mitigate algorithmic bias in AI systems that consistently predict a favorable outcome for a certain group, but improve overall accuracy by mitigating class imbalance within data that leads to a bias towards the majority class. We demonstrate the effectiveness of our technique on real datasets using a variety of fairness metrics.
BibTeX
@misc{zhou2021improving,
  author = {Yan Zhou and Murat Kantarcioglu and Chris Clifton},
  title = {{Improving Fairness of AI Systems with Lossless De-biasing}},
  year = {2021},
  eprint = {2105.04534},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2105.04534},
}
Figure