Title: Quantifying (Un)Fairness
Presented By: Mohamed Abdalla, University of Toronto

In the machine learning fairness literature, the majority of fairness definitions are formalized for the binary case. This binary formalization allows for simple hypothetical demonstrations of the fairness definitions and provides researchers the ability to prove theorems. However, in clinical settings, the binary case is often too simple. As a consequence, we must expand machine learning definitions of fairness beyond these binary formulations. In this work, we analyze different ways of expanding fairness definitions beyond the binary case, highlighting edge cases where such expansions do not work as expected. We perform empirical analysis on a clinical task to assess how likely each edge case is for non-binary expansion, and consider the clinical feasibility and ramifications of such decisions.

For more information, visit the online event page.