Fairness-Aware Learning Algorithms for Responsible Decision Making

Abstract

Fairness-aware learning algorithms seek to produce predictive models that satisfy formal fairness criteria -- including demographic parity, equalised odds, and counterfactual fairness -- while preserving predictive utility. Despite considerable algorithmic progress, the practical deployment of fairness-aware learning in real-world decision systems faces three persistent challenges: the multiplicity problem (no single algorithm simultaneously satisfies all fairness criteria in all contexts), the accuracy-fairness trade-off (fairness constraints typically reduce predictive accuracy), and the context-dependency problem (optimal algorithm selection depends on domain-specific ethical and legal requirements). This paper presents a systematic comparative evaluation of eight fairness-aware learning algorithms across four deployment domains -- credit scoring, clinical triage, recidivism prediction, and university admissions -- using a multi-criteria evaluation framework assessing five fairness metrics, two accuracy metrics, and three robustness metrics across n = 16 benchmark datasets. Results demonstrate that no algorithm dominates across all criteria: adversarial debiasing achieves the best mean demographic parity difference (DPD = 0.031, SD = 0.008) but the highest accuracy cost (mean AUC reduction = 4.7%), while prejudice remover achieves moderate fairness improvement with minimal accuracy cost. A Pareto-optimal algorithm selection framework is introduced, mapping domain-specific fairness obligation profiles to optimal algorithm choices. The study contributes a replicable benchmark methodology, a fairness-accuracy-robustness evaluation dashboard, and a decision support tool for algorithm selection aligned with EU AI Act non-discrimination requirements.