Date: Wednesday, June 15th, 2022
9:00 am – 10:00 am Pacific Time
12:00 pm – 1:00 pm Eastern Time
Location: Weekly Seminar, Zoom
Title: Secrets, Adversaries, Incentives, and Composition in Algorithmic Fairness
A central goal of algorithmic fairness is to build systems with fairness properties that compose gracefully. A major effort towards this goal in fair machine learning has been the development of fair representations which guarantee demographic parity under sequential composition by removing group membership information from the data (i.e. by imposing a demographic secrecy constraint). This approach models all data consumers as utterly malicious adversaries whose sole objective is to be as unfair as possible (i.e. maximally violate demographic parity)—any other possible objective (i.e. incentive) that a data consumer may have is dismissed and ignored. In this talk, I describe joint work with Augustin Chaintreau, in which we elucidate limitations of demographically secret fair representations and propose a fresh approach to potentially overcome them by incorporating information about parties’ incentives into fairness interventions. We show that in a stylized model, it is possible to relax demographic secrecy to obtain incentive-compatible representations, where rational parties obtain exponentially greater utilities vis-à-vis any demographically secret representation and satisfy demographic parity. These substantial gains are recovered not from the well-known cost of fairness, but rather from a cost of demographic secrecy which we formalize and quantify for the first time. We further show that the sequential composition property of demographically secret representations is not robust to aggregation. Our results contribute to further understanding the challenges of fair composition while simultaneously suggesting that incentives may be an important and flexible tool for addressing or even overcoming those challenges.
Roland Maio is a fourth year Computer Science PhD student at Columbia University advised by Augustin Chaintreau. Roland works on algorithmic fairness and CS ethics. His work has been supported by an NSF fellowship.