Criteria for algorithmic fairness metric selection under different supervised classification scenarios

Breger, Clothilde

Criteria for algorithmic fairness metric selection under different supervised classification scenarios

License

Citation

Breger, C. Criteria for algorithmic fairness metric selection under different supervised classification scenarios. 2020. handle: http://hdl.handle.net/10230/46359

This citation was generated automatically.

Abstract

The research community, (supra-)national institutions, and regular users have noticed that Artificial Intelligence and Machine Learning algorithms can amplify existing inequity between groups. One way to limit this is to use group fairness metrics to measure inequity, optimise and select models. However, there are many different group fairness metrics. Here I combined a clustering of metrics (as done by Friedler et al. in their 2019 paper "A comparative study of fairness-enhancing interventions in machine learning" and by Miron et al. in their 2020 paper "Addressing multiple metrics of group fairness in data-driven decision making") and expert-driven recommendations (from a case study by Rodolfa et al., published in 2020: "Case study: Predictive Fairness to Reduce Misdemeanor Recidivism Through Social Service Interventions") to select fairness metrics. Although this clustering was not consistent, it enabled fairness metric selection and fostered general recommendations on the matter: an algorithm designer should extensively study their algorithm’s application context and explicitly justify their choices relative to fairness. So long as there is no absolute guide to metric selection, this should help nourish an ongoing and context-specific discussion on algorithmic fairness, within and outside of the research community.