Towards Practical Notions of Individual Fairness

Thursday, June 20, 2019 - 10:00am - 10:50am
Lind 305
Aaron Roth (University of Pennsylvania)
The field of fairness in machine learning has yet to settle on definitions. At a high level, there are (at least) two approaches: individual, and statistical definitions of fairness. Statistical definitions of fairness ask that some statistic of the classifier (like error rate or false positive rate) be equalized across some set of groups. While this kind of constraint is easy to check and to achieve, it provides only a very limited promise to individuals. Individual definitions of fairness, on the other hand, aspire to stronger semantic meaning, providing promises to particular individuals. Unfortunately, the definitions of individual fairness proposed to date make assumptions of different sorts that seem to be a substantial barrier to practical use. In this work, we propose two variants of existing notions of individual fairness that can circumvent the need to make strong assumptions. They provide individual-level guarantees while still being actionable in practice.