Fairness-Aware Programming

Abstract

Increasingly, programming tasks involve automating and deploying sensitive decision-making processes that may have adverse impacts on individuals or groups of people. The issue of fairness in automated decision-making has thus become a major problem, attracting interdisciplinary attention. In this work, we aim to make fairness a first-class concern in programming. Specifically, we propose fairness-aware programming, where programmers can state fairness expectations natively in their code, and have a runtime system monitor decision-making and report violations of fairness. We present a rich and general specification language that allows a programmer to specify a range of fairness definitions from the literature, as well as others. As the decision-making program executes, the runtime maintains statistics on the decisions made and incrementally checks whether the fairness definitions have been violated, reporting such violations to the developer. The advantages of this approach are two fold: (i) Enabling declarative mathematical specifications of fairness in the programming language simplifies the process of checking fairness, as the programmer does not have to write ad hoc code for maintaining statistics. (ii) Compared to existing techniques for checking and ensuring fairness, our approach monitors a decision-making program in the wild, which may be running on a distribution that is unlike the dataset on which a classifier was trained and tested. We describe an implementation of our proposed methodology as a library in the Python programming language and illustrate its use on case studies from the algorithmic fairness literature.