FairSelect Evaluates Algorithmic Fairness Across Subgroups

As machine learning models become more prevalent, ensuring algorithmic fairness and mitigating bias is paramount. However, most fairness methods are evaluated in isolation and along single demographic axes, which fails to capture the complexities of bias arising across intersectional subgroups and different stages of the modeling process. Researchers have introduced FairSelect, a comprehensive toolkit designed to systematically evaluate fairness mitigation strategies. This framework allows for the assessment of individual and combined fairness interventions across preprocessing, inprocessing, and postprocessing stages, supporting multiple model architectures and intersectional subgroup analysis. Validation using synthetic clinical datasets and a real-world stroke risk prediction task revealed that fairness interventions interact in non-additive and context-dependent ways. While targeted methods generally reduced intended disparities, combined strategies often yielded larger average improvements with modest utility tradeoffs. Critically, some combinations improved both fairness and predictive performance, while others were ineffective or counterproductive, highlighting the need for systematic evaluation.