Fairness Through the Lens of Proportional Equality

Today, automated algorithms, such as machine learning classifiers, are playing an increasingly pivotal role in important societal decisions such as hiring, loan allocation, and criminal risk assessment. This motivates the need to probe the outcomes of a prediction model for discriminatory traits towards specific groups of individuals. In this context, one of the crucial challenges is to formally define a satisfactory notion of fairness. Our contribution in this paper is to formalize Proportional Equality (PE) as a fairness notion. We additionally show that it is a more appropriate criterion than the existing popular notion called Disparate Impact (DI), which is used for evaluating the fairness of a classifier's outcomes.