Title: From Soft Classifiers to Hard Decisions

Presented By: Govind Ramnaryan, MIT

Abstract:

In recent years, it has become clear that many important decisions in people's lives will be made with significant input from machine learning algorithms. Thus, researchers have begun to explore what it means for a decision-making algorithm to be fair. Unfortunately, there are many such definitions, and they are often incompatible with each other in strong ways, seemingly leaving little room for satisfying algorithms [Kleinberg et al 2016, Chouldechova 2017].

In this work, we study a more specific class of algorithms that is commonly used in practice. First, we show additional impossibility results for such algorithms. Finally, we conclude on a more positive note by additionally suggesting a work-around: deferring decision-making, which was originally introduced to fairness literature by Madras et al 2017.

Joint work with Ran Canetti, Aloni Cohen, Nishanth Dikkala, Sarah Scheffler, and Adam Smith.