FairPOT: Balancing AUC Performance and Fairness with Proportional Optimal Transport.

Area of Science:

• Machine Learning
• Algorithmic Fairness
• Data Science

Background:

• Fairness metrics using Area Under the Curve (AUC) are crucial in high-stakes fields.
• Enforcing strict fairness often compromises AUC performance, especially with risk scores.
• Existing methods struggle to balance fairness and utility effectively.

Purpose of the Study:

• To introduce Fair Proportional Optimal Transport (FairPOT), a novel post-processing framework for model-agnostic fairness.
• To enable tunable trade-offs between reducing AUC disparities and maintaining overall AUC performance.
• To extend fairness interventions to the partial AUC setting, focusing on high-risk predictions.

Main Methods:

• FairPOT employs optimal transport to align risk score distributions across groups.
• It selectively transforms a controllable proportion (top-λ quantile) of scores in disadvantaged groups.
• The framework is extended for partial AUC scenarios, concentrating on critical prediction regions.

Main Results:

• FairPOT consistently outperforms existing post-processing techniques in both global and partial AUC settings.
• It achieves improved fairness with minimal AUC degradation, sometimes even enhancing utility.
• Experiments on diverse datasets validate FairPOT's effectiveness and adaptability.

Conclusions:

• FairPOT offers a computationally efficient and adaptable solution for achieving algorithmic fairness.
• The method provides a tunable approach to balance fairness and performance in risk score-based systems.
• FairPOT is a promising tool for real-world deployment in sensitive applications.