Context-conscious fairness in using machine learning to make decisions

Abstract

The increasing adoption of machine learning to inform decisions in employment, pricing, and criminal justice has raised concerns that algorithms may perpetuate historical and societal discrimination. Academics have responded by introducing numerous definitions of "fairness" with corresponding mathematical formalisations, proposed as one-size-fits-all, universal conditions. This paper will explore three of the definitions and demonstrate their embedded ethical values and contextual limitations, using credit risk evaluation as an example use case. I will propose a new approach - context-conscious fairness - that takes into account two main trade-offs: between aggregate benefit and inequity and between accuracy and interpretability. Fairness is not a notion with absolute and binary measurement; the target outcomes and their trade-offs must be specified with respect to the relevant domain context.