Fairness in Machine Learning

Important decisions such as police deployment, loan approvals, hiring, and parole from incarceration are beginning to be made by machine learning algorithms. This has lent urgency to the question of whether these algorithms are fair. A recent well-publicized analysis of recidivism risk predictions studied NorthPointe’s proprietary COMPAS algorithm used in Broward County, Florida. Although the algorithm was well-calibrated across both black and white populations, its false-positive rate was significantly higher amongst black defendants. In other words, black defendants who did not go on to recidivate were substantially more likely to be labelled as “high risk” by the COMPAS algorithm than were white defendants who did not go on to recidivate. This means that the costs of the (inevitable) inaccuracy of the COMPAS algorithm accrued disproportionately to the black population.

“Fairness” is a challenging goal to precisely define and achieve. There is an extensive literature in philosophy, ethics, law, and the social sciences. Drawing on this literature, we seek to find quantitative and computationally tractable definitions. Foundational results imply that we cannot have it all: technical discussions of fairness desiderata must be oriented towards quantifying tradeoffs between different goals. We ask:

1. What are reasonable precise definitions of fairness? Are they mutually exclusive, or can they be simultaneously achieved?
2. At what cost is fairness achieved in terms of other desiderata? How much must we lose in classification accuracy in order to be “fair”?
3. Can regulators incentivize decision makers to be fair without needing to explicitly constrain their actions? If so, at what (financial) cost?
4. Can we prove that fairness is impossible to achieve under certain definitions? If so, what can we do as an alternative? Can we limit the number of  “unfair” decisions or otherwise mitigate the situation?

Please see our Penn News Feature for more details!