Mitigating Bias and Improving Fairness in Predictive Machine Learning Models for Financial Risk Assessment
Digital Transformation and Technology Dynamics · 29.11.2021 · 1
DOI: 10.69013/xx4349mz ↗Kurz gefasst
Fachartikel in der Zeitschrift Digital Transformation and Technology Dynamics, online erschienen am 29.11.2021. Verfasst von MD Shahedul Islam Khan und 1 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben.
Abstract (Originalsprache)
Machine learning (ML) systems increasingly mediate access to credit, insurance, and other financial services, yet these systems routinely encode and amplify historical patterns of discrimination against protected demographic groups. The tension between predictive accuracy, the primary driver of lender profitability, and demographic fairness constitutes an unresolved socio-technical problem with direct regulatory and welfare consequences. This paper develops a formal mathematical treatment of the principal group-fairness criteria, including demographic parity, equalized odds, equal opportunity, and predictive parity/disparate impact, as applied to financial risk scoring, and proves their pairwise incompatibility under realistic base-rate conditions. Building on this foundation, we propose a Hybrid Adversarial-Recourse Debiasing Framework (HARD-F) that couples in-processing adversarial representation learning with a post-processing fairness-aware recourse and threshold-calibration engine, explicitly designed for deployment within enterprise MLOps credit-risk pipelines. Empirical evaluation on the German Credit, Taiwanese Credit Card Default, and a synthetic high-volume mortgage-origination dataset shows that HARD-F reduces disparate impact ratio deviation from the 80% threshold by 61–74% relative to unconstrained baselines, including Logistic Regression, XGBoost, Random Forest, and a feed-forward neural network, while sacrificing only 1.8–3.4 AUC-ROC points and dominating five established mitigation baselines on the empirical Pareto frontier. SHAP-based explainability analysis reveals substantial pre-to-post-mitigation attribution shifts away from geography-correlated proxy variables toward income-stability features. The results offer risk officers, model validators, and regulators a quantitatively grounded, auditable pathway toward ECOA- and EU-AI-Act-compliant credit models without prohibitive loss of discriminative power and identify continuous drift monitoring and federated fair learning as priority directions for future work.
Zitieren
APA
Khan, M. S. I., & Shishir, M. S. H. (2021). Mitigating Bias and Improving Fairness in Predictive Machine Learning Models for Financial Risk Assessment. Digital Transformation and Technology Dynamics, 1. https://doi.org/10.69013/xx4349mz
Vancouver
Khan MSI, Shishir MSH. Mitigating Bias and Improving Fairness in Predictive Machine Learning Models for Financial Risk Assessment. Digital Transformation and Technology Dynamics. 2021:1. doi:10.69013/xx4349mz
BibTeX
@article{Khan2021,
author = {Khan, MD Shahedul Islam and Shishir, Md Shahadat Hossain},
title = {Mitigating Bias and Improving Fairness in Predictive Machine Learning Models for Financial Risk Assessment},
journal = {Digital Transformation and Technology Dynamics},
year = {2021},
pages = {1},
doi = {10.69013/xx4349mz}
}
EUVIMED