Computational Approaches to Detect and Mitigate Algorithmic Bias

Abstract

Algorithmic bias -- the systematic and unjustifiable differential treatment of individuals based on protected characteristics such as race, gender, age, or disability -- pervades deployed machine learning systems and constitutes one of the primary technical challenges of responsible AI. While individual detection metrics and mitigation algorithms have been extensively studied, a comprehensive computational framework integrating bias detection across multiple fairness criteria, mitigation method selection, and post-mitigation verification remains absent. This paper proposes the Bias Detection and Mitigation Suite (BDMS), a modular computational framework comprising four integrated components: a Multi-Metric Bias Scanner (MMBS) computing fifteen fairness metrics across all protected attributes and intersectional subgroups; an Automated Mitigation Recommender (AMR) that selects optimal mitigation strategies using a constraint-satisfaction algorithm aligned with domain-specific fairness obligation profiles; a Mitigation Effectiveness Verifier (MEV) that evaluates post-mitigation outcomes against pre- specified fairness targets; and a Bias Audit Reporter (BAR) generating EU AI Act-compliant audit documentation. The BDMS is evaluated on 20 benchmark datasets across five domains and compared against three existing bias toolkits. Results demonstrate that BDMS achieves a 23.7% higher mean intersectional fairness improvement than single- criterion tools (p < 0.001) and a 31.4% lower false-negative rate for bias detection (detecting previously undetected bias in 8 of 20 datasets). The AMR recommendation accuracy is 84.6% (SD = 6.2%) against expert-panel ground truth. The BDMS contributes a replicable open-source computational toolkit, a 15-metric fairness evaluation standard, and an intersectional bias detection methodology for responsible AI practice.