Framework

Algorithmic fairness metrics

The question it answers

What does 'fair' mean in a machine-learning model, and why can't we have everything at once?

Origin
Dwork et al., 2012; Kleinberg et al., 2017; Binns, 2018
Diagram type
Comparison table
Level of analysis
Team
Topics
Data & analyticsAI for leaders
Magnifying lens over a line chart tile, icon for data and analytics

Diagram

Algorithmic fairness metrics

CriterionMeaningTension / limit
Individual fairnessSimilar people are treated similarlyNeeds a definition of 'similar'
Demographic paritySelection rates match population sharesHard with small samples and intersecting traits
Equal opportunityEqual true-positive rates across groupsImpossibility result: not every metric can be met at once
Error-rate balanceEqual error rates given the true outcomeImpossibility result: not every metric can be met at once
Counterfactual fairnessThe outcome would not change if the person belonged to another groupImpossibility result: not every metric can be met at once

Sources

References

Listed A–Z by author surname, Harvard style. A source with a DOI links to https://doi.org/…; one without links to its page or says plainly that it has no DOI.

  1. Binns, R. (2018) ‘Fairness in machine learning: lessons from political philosophy’, in Proceedings of Machine Learning Research. pp. 149–159.No DOIproceedings.mlr.pressSource page
  2. Dwork, C. et al. (2012) ‘Fairness through awareness’, in Proceedings of the 3rd Innovations in Theoretical Computer Science Conference. New York: ACM, pp. 214–226.doi:10.1145/2090236.2090255Source page
  3. Kleinberg, J., Mullainathan, S. and Raghavan, M. (2017) ‘Inherent trade-offs in the fair determination of risk scores’, in 8th Innovations in Theoretical Computer Science Conference (ITCS 2017), LIPIcs 67. pp. 43:1–43:23.doi:10.4230/LIPIcs.ITCS.2017.43Source page