Monday, November 19, 2018
Wegmans Hall 1400
Professor Ran Canetti
From Soft Classifers to Hard Decisions: How Fair Can We Be?
As society is charging ahead with the prevalent use of (often opaque) algorithms in decision-making processes regarding people, it has become critical to study the societal implications of such algorithms. This study is fascinating, in that it forces us to mathematically formulate and quantify concepts such as fairness and social justice, which have traditionally been thought of in qualitative and subjective terms. In particular, this study is teasing out inherent inconsistencies in traditional thinking and helping us clarify our social goals.
We consider one slice of this study: The need to balance the errors across societal groups in algorithmic decision-making based on imperfect information. Here some widely used statistical measures error-balance turn out to be contradictory (Kleinberg et al. '16, Chouldechova '17). We propose a way to - at least partially - evade this bound. Specifically, we observe that if the algorithm is allowed to `defer' on some decisions (that is, the algorithm can avoid making some decisions and hand them off to a separate process which is presumably more expensive and more precise), then the errors in the non-deferred decisions will be balanced across the considered societal groups by all statistical measures currently under consideration. We further demonstrate concrete ways for using deferrals in the natural context of post-processing ``soft’’ classifiers that output a numerical score into a binary decision. We evaluate our algorithms using the COMPAS data set from 2016 and consider potential societal implications of deferrals.