An End-to-End Responsible AI Framework for Bias Detection, Multi-Metric Fairness Evaluation, and Real-Time Drift Monitoring in Machine Learning Classification Models

This paper presented a modular Responsible AI framework for bias detection, mitigation, and continuous fairness monitoring in supervised machine learning classification models. The framework addresses four gaps in prior work: (1) a unified end-to-end pipeline absent from existing toolkits; (2) a novel Predictive Equality Difference metric and a principled avg\_bias threshold calibrated against NIST AI RMF guidance; (3) fairness drift detection validated on real temporal data streams (Taiwan Credit Default and MIMIC-III); and (4) SHAP-attributed explainability producing actionable proxy variable identification and regulatory-aligned audit reports. Experimental evaluation across five benchmark datasets demonstrated that the multi-stage mitigation pipeline achieves fairness metric reductions of 76–82% with less than 3% accuracy degradation. Real-world temporal drift detection achieved sub-period sensitivity on two independent datasets. SHAP attribution identified actionable proxy variables whose removal yields additional fairness improvements beyond post-processing. Future work will investigate: (1) federated fairness monitoring for distributed deployment contexts; (2) intersectionality-aware metrics using multi-dimensional Wasserstein distance; (3) application to generative AI models; and (4) integration with the EU AI Act conformity assessment process through automated documentation generation. The framework is released as open-source software to support fairness-aware practices in real-world AI deployment.

[1]  N. Mehrabi et al., "A survey on bias and fairness in machine learning," ACM Computing Surveys, vol. 54, no. 6, pp. 1-35, 2021. [2]  D. Pessach and E. Shmueli, "A review on fairness in machine learning," ACM Computing Surveys, vol. 55, no. 3, pp. 1-44, 2022. [3]  B. Giovanola and S. Tiribelli, "Beyond bias and discrimination: redefining the AI ethics principle of fairness in healthcare ML algorithms," AI & Society, vol. 38, no. 2, pp. 549-563, 2023. [4]  D. B. Patel, "Ethical AI: Addressing bias and fairness in machine learning models," J. Computer Science and Technology Studies, vol. 3, no. 1, pp. 13-17, 2021. [5]  F. Osasona et al., "Reviewing the ethical implications of AI in decision making processes," Int. J. Management & Entrepreneurship Research, vol. 6, no. 2, pp. 322-335, 2024. [6]  J. Angwin et al., "Machine bias," ProPublica, May 2016. [7]  J. Dastin, "Amazon scraps secret AI recruiting tool that showed bias against women," Reuters, Oct. 2018. [8]  Z. Obermeyer et al., "Dissecting racial bias in an algorithm used to manage the health of populations," Science, vol. 366, no. 6464, pp. 447-453, 2019. [9]  J. K. Bahangulu and L. Owusu-Berko, "Algorithmic bias, data ethics, and governance," World J. Advanced Research and Reviews, vol. 25, no. 2, pp. 1746-1763, 2025. [10]  M. G. Hanna et al., "Ethical and bias considerations in artificial intelligence/machine learning," Modern Pathology, vol. 38, p. 100686, 2025.