Sorelle A. Friedler, Carlos Scheidegger, Suresh Venkatasubramanian at Communications of the ACM: “Automated decision-making systems (often machine learning-based) now commonly determine criminal sentences, hiring choices, and loan applications. This widespread deployment is concerning, since these systems have the potential to discriminate against people based on their demographic characteristics. Current sentencing risk assessments are racially biased, and job advertisements discriminate on gender. These concerns have led to an explosive growth in fairness-aware machine learning, a field that aims to enable algorithmic systems that are fair by design.
To design fair systems, we must agree precisely on what it means to be fair. One such definition is individual fairness: individuals who are similar (with respect to some task) should be treated similarly (with respect to that task). Simultaneously, a different definition states that demographic groups should, on the whole, receive similar decisions. This group fairness definition is inspired by civil rights law in the U.S. and U.K. Other definitions state that fair systems should err evenly across demographic groups. Many of these definitions have been incorporated into machine learning pipelines.
In this article, we introduce a framework for understanding these different definitions of fairness and how they relate to each other. Crucially, our framework shows these definitions and their implementations correspond to different axiomatic beliefs about the world. We present two such worldviews and will show they are fundamentally incompatible. First, one can believe the observation processes that generate data for machine learning are structurally biased. This belief provides a justification for seeking non-discrimination. When one believes that demographic groups are, on the whole, fundamentally similar, group fairness mechanisms successfully guarantee the top-level goal of non-discrimination: similar groups receiving similar treatment. Alternatively, one can assume the observed data generally reflects the true underlying reality about differences between people. These worldviews are in conflict; a single algorithm cannot satisfy either definition of fairness under both worldviews. Thus, researchers and practitioners ought to be intentional and explicit about world-views and value assumptions: the systems they design will always encode some belief about the world….(More)”.