On the (im)possibility of fairness

• DOI: 10.48550/arxiv.1609.07236
• Publication Date: 2016-01-01
• Authors: Sorelle A. Friedler; Carlos Scheidegger; Suresh Venkatasub...
• Keywords: Unobservable; Decision process

What does it mean for an algorithm to be fair? Different papers use different notions of algorithmic fairness, and although these appear internally consistent, they also seem mutually incompatible. We present a mathematical setting in which the distinctions previous can made formal. In addition characterizing spaces inputs (the "observed" space) outputs "decision" space), we introduce notion construct space: space that captures unobservable, but meaningful variables prediction. show order prove desirable properties entire decision-making process, mechanisms fairness require assumptions about nature mapping from decision space. The results this paper imply future treatments should more explicitly state relationship between constructs observations.