This paper presents a context in which there is growing interest in fairness metrics using the area under the receiver operating characteristic curve (AUC) in high-stakes fields such as healthcare, finance, and criminal justice. In these fields, fairness is often assessed based on risk scores rather than binary outcomes, and applying strict fairness standards can significantly degrade AUC performance. To address this issue, we propose Fair Proportional Optimal Transport (FairPOT), a novel model-independent postprocessing framework that strategically aligns the distribution of risk scores across different groups using optimal transport, but selectively transforms a controllable proportion of scores within the disadvantaged group (the top lambda quantile). By varying lambda, the proposed method allows for a tunable trade-off between reducing AUC imbalance and maintaining overall AUC performance. Furthermore, we extend FairPOT to partial AUC settings, allowing for targeted fairness interventions to be targeted at the highest-risk regions. Extensive experiments on synthetic, public, and clinical data demonstrate that FairPOT consistently outperforms existing postprocessing techniques in both global and partial AUC scenarios, often improving fairness with only a slight decrease in AUC or a positive improvement in utility. FairPOT's computational efficiency and practical adaptability make it a promising solution for real-world deployment.
