Fairness-Aware Feature Attribution for Credit Scoring: A Causal Path Decomposition Approach

Abstract

Credit scoring algorithms increasingly influence financial inclusion outcomes, yet traditional approaches often encode discriminatory patterns that disadvantage protected groups. This paper presents a fairness-aware feature attribution framework that leverages causal path decomposition to distinguish legitimate predictive factors from discriminatory proxies in credit assessment. The proposed approach integrates Shapley Additive Explanations with causal directed acyclic graphs to quantify the fair and unfair contributions of each feature to credit decisions. Experimental validation on two benchmark datasets demonstrates that the framework improves the disparate impact (DI) ratio while retaining 94.2% of baseline predictive performance (AUC). The causal feature filtering mechanism identifies features whose contributions are dominated by proxy-discrimination effects. It mitigates such influence through targeted feature filtering, enabling financial institutions to develop credit-scoring algorithms that satisfy both regulatory compliance requirements and business performance objectives. This research provides practical guidance on integrating alternative data sources while preserving fairness constraints, thereby directly supporting the Consumer Financial Protection Bureau's algorithmic discrimination-prevention goals and the Community Reinvestment Act's financial-inclusion mandates.