Scholarlink

<p>As AI and ML applications become more common in public organisations, technical processes are needed to audit model behaviour, document limitations, and mitigate measurable disparities before deployment. This study presents a methodological framework for fairness auditing and bias mitigation in supervised classification, with emphasis on error disparities across groups rather than a complete solution to AI ethics. The framework is illustrated with two experiments based on data from Chile’s Public Criminal Defence Office (DPP), where the task is to predict whether a criminal case has a favourable or unfavourable outcome for the defendant. Each experiment compares a base model with an improved model that uses reweighting, class weighting, threshold optimisation, interpretability analysis, and documentation. Fairness is operationalised through group-level error metrics, especially false negative rate (FNR), false omission rate (FOR), and true positive rate (TPR), using Region as a geographic proxy attribute. The results show that mitigation can reduce selected disparities, but also reveal a clear fairness–performance trade-off. In the drug trafficking experiment, the manuscript-level confusion matrices show a reduction in FNR from 40.54 to 31.63%, with an increase in FPR from 0.13 to 39.25%. Statistical comparisons are computed from the confusion-matrix counts reported in this manuscript and are therefore reported as aggregate two-proportion comparisons rather than paired McNemar tests. The framework should therefore be read as a technical audit and mitigation procedure, not as evidence that the resulting models are suitable for operational legal decision-making.</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

Fairness and transparency in ML: a methodological framework

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月25日

查看全文
  • Nelson Salazar Valdebenito,
  • Gonzalo A. Ruz,
  • Reinel Tabares-Soto

摘要

As AI and ML applications become more common in public organisations, technical processes are needed to audit model behaviour, document limitations, and mitigate measurable disparities before deployment. This study presents a methodological framework for fairness auditing and bias mitigation in supervised classification, with emphasis on error disparities across groups rather than a complete solution to AI ethics. The framework is illustrated with two experiments based on data from Chile’s Public Criminal Defence Office (DPP), where the task is to predict whether a criminal case has a favourable or unfavourable outcome for the defendant. Each experiment compares a base model with an improved model that uses reweighting, class weighting, threshold optimisation, interpretability analysis, and documentation. Fairness is operationalised through group-level error metrics, especially false negative rate (FNR), false omission rate (FOR), and true positive rate (TPR), using Region as a geographic proxy attribute. The results show that mitigation can reduce selected disparities, but also reveal a clear fairness–performance trade-off. In the drug trafficking experiment, the manuscript-level confusion matrices show a reduction in FNR from 40.54 to 31.63%, with an increase in FPR from 0.13 to 39.25%. Statistical comparisons are computed from the confusion-matrix counts reported in this manuscript and are therefore reported as aggregate two-proportion comparisons rather than paired McNemar tests. The framework should therefore be read as a technical audit and mitigation procedure, not as evidence that the resulting models are suitable for operational legal decision-making.