Sector-Specific Fairness in AI: A Quasi-Optimization Weight Learning Framework for Critical Industries

Abstract

The "one-size-fits-all" paradigm for AI fairness poorly accommodates industrial heterogeneity across healthcare, finance, education, and manufacturing. Uniform fairness enforcement may increase medical misdiagnosis by 12% and distort credit pricing by 18%. This paper proposes a Quasi-Optimization Weight Learning (QOWL) framework to learn sector-specific fairness weights via bilevel constrained optimization. Grounded in bounded rationality, QOWL yields interpretable, deployable suboptimal solutions instead of strict global optimality. We build a real-world AI fairness dataset of 875 incident records (2015–2025). Experiments reveal distinct sectoral fairness patterns: manufacturing surprisingly prioritizes Individual Fairness (weight = 0.35), as violations cause immediate production halts and economic losses. Compared with uniform weighting and competitive baselines, QOWL reduces fairness violations by 27%–44% and improves Pareto efficiency to 0.79–0.84, with statistically significant gains (p<0.01 for healthcare/finance, p<0.05 for education/manufacturing). Our results are mapped to Shenzhen AI industry governance requirements, including AI testing, ethical safety risk assessment, privacy protection, and algorithm explanation obligations, to provide quantitative, sector-differentiated benchmarks for practical AI regulation.