Scholarlink

<p>As algorithmic systems increasingly mediate access to opportunity, justice, and resources, ensuring their fairness is both a technical and ethical imperative. This paper examines the ethical and technical dimensions of fairness-enhancing interventions in algorithmic decision-making systems through an intersectional lens. While prior work often addresses fairness in isolated stages or relies on aggregate group metrics, we assess integrated mitigation strategies that span pre-, in-, and post-processing, using disaggregated subgroup analysis. Drawing on two benchmark datasets (COMPAS (recidivism) and Adult Income (Census Data), we evaluate 27 model configurations across four fairness metrics (Statistical Parity Difference, Disparate Impact, Equal Opportunity Difference, and Predictive Equality Difference) and predictive accuracy. Our findings show that although multi-stage interventions can improve overall fairness with minimal accuracy loss, aggregate metrics frequently conceal systemic harms toward marginalized subgroups, particularly Black women. We also introduce enhanced methods (DIR+ and AD+) tailored for intersectional contexts, where individuals belong to multiple overlapping protected groups (e.g., race <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> gender), and disparities may emerge not just along single dimensions, but at their intersections. The results highlight how the prevailing fairness frameworks, absent intersectional analysis, risk ethical blind spots, and reinforce structural inequities. We argue that intersectionally disaggregated auditing is not only methodologically essential but ethically non-negotiable for achieving just algorithmic systems. This study provides empirical evidence and normative guidance for designing fairness-aware AI systems that are responsive to compound and context-specific forms of bias.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Beyond aggregate fairness: intersectional auditing across the AI fairness pipeline

  • 期刊论文

  • 发表日期：

    2026年1月12日

查看全文
  • Michael Mayowa Farayola,
  • Malika Bendechache,
  • Takfarinas Saber,
  • Regina Connolly,
  • Irina Tal

摘要

As algorithmic systems increasingly mediate access to opportunity, justice, and resources, ensuring their fairness is both a technical and ethical imperative. This paper examines the ethical and technical dimensions of fairness-enhancing interventions in algorithmic decision-making systems through an intersectional lens. While prior work often addresses fairness in isolated stages or relies on aggregate group metrics, we assess integrated mitigation strategies that span pre-, in-, and post-processing, using disaggregated subgroup analysis. Drawing on two benchmark datasets (COMPAS (recidivism) and Adult Income (Census Data), we evaluate 27 model configurations across four fairness metrics (Statistical Parity Difference, Disparate Impact, Equal Opportunity Difference, and Predictive Equality Difference) and predictive accuracy. Our findings show that although multi-stage interventions can improve overall fairness with minimal accuracy loss, aggregate metrics frequently conceal systemic harms toward marginalized subgroups, particularly Black women. We also introduce enhanced methods (DIR+ and AD+) tailored for intersectional contexts, where individuals belong to multiple overlapping protected groups (e.g., race \(\times \) gender), and disparities may emerge not just along single dimensions, but at their intersections. The results highlight how the prevailing fairness frameworks, absent intersectional analysis, risk ethical blind spots, and reinforce structural inequities. We argue that intersectionally disaggregated auditing is not only methodologically essential but ethically non-negotiable for achieving just algorithmic systems. This study provides empirical evidence and normative guidance for designing fairness-aware AI systems that are responsive to compound and context-specific forms of bias.