- Moving Beyond Observational Notions of Fairness
- Project Year:
2018
- REU Student (s):
Michael Yang | Minerva Schools at Keck Graduate Institute CA
- Student 1 Institution:
Minerva Schools at Keck Graduate Institute
- Project Mentor:
Anand Sarwate
- Project Mentor Area:
Electrical and Computer Engineering
- Project Abstract:
Computer scientists have already unleashed a bevy of technical definitions of fairness. It can be confusing for a newcomer to the field to make sense of the various definitions, their implications, and their shortcomings. In this article, we review previous and contemporary definitions of fairness. Earlier work in the field generally focuses on assessing and learning classifiers with respect to so-called observational notions of fairness. However, more recent work pointed out inherent limitations in these fairness definitions. That a classifier is fair with respect to a fixed notion of fairness is often not sufficient to address all intuitive unfairness in its predictions or the situation in which the algorithm is deployed. The upshot of this review is that "fair" algorithms must incorporate more information into the decision that is assessed merely by observational notions of fairness. Our review surveys three kinds of additional information: the causal structure between variables, the intersection of arbitrarily many protected categories, and the long-term effects of different predictions.
