Fairness-Aware Algorithmic Recourse for Credit Scoring Models

Abstract

Algorithmic recourse offers individuals affected by automated credit decisions actionable pathways to reverse unfavorable outcomes. However, recourse opportunities are often inequitably distributed across demographic groups, imposing systematically higher improvement burdens on protected populations a disparity that conventional fairness diagnostics fail to detect. This paper presents an integrated fairness-aware recourse framework for credit scoring that unites three key innovations: (1) pairwise feature preference elicitation to measure individualized recourse costs, (2) coverage-validity-aware surrogate modeling to generate recourses robust to future model retraining, and (3) ranked group-level fairness constraints that ensure equitable recourse distribution across protected groups. Empirical evaluation on the German Credit Dataset demonstrates that fairness-unaware recourse generation imposes gender and age cost disparities of 17.6% and 21.0%, respectively. Our framework reduces these gaps to 2.3% and 3.0% while maintaining a 91.8% recourse validity rate and preserving 89.4% of recourse validity across model retraining cycles a substantial improvement over the baseline's 52.8% preservation rate. Feature importance and causal counterfactual analyses further reveal that the framework shifts recourse recommendations toward more equitably distributed features and that a majority of the observed unfairness is attributable to the recourse generation mechanism itself. These findings underscore that algorithmic recourse fairness is both measurable and achievable without compromising practical utility. The framework offers financial institutions a transparent, auditable approach to generating robust and equitable recourses, contributing to responsible AI deployment in high-stakes lending decisions.