Matching code and law: achieving algorithmic fairness with optimal transport

Abstract

Increasingly, discrimination by algorithms is perceived as a societal and legal problem. As a response, a number of criteria for implementing algorithmic fairness in machine learning have been developed in the literature. This paper proposes the Continuous Fairness Algorithm (CFAθ ) which enables a continuous interpolation between different fairness definitions. More specifically, we make three main contributions to the existing literature. First, our approach allows the decision maker to continuously vary between specific concepts of individual and group fairness. As a consequence, the algorithm enables the decision maker to adopt intermediate "worldviews" on the degree of discrimination encoded in algorithmic processes, adding nuance to the extreme cases of "we're all equal" (WAE) and "what you see is what you get" (WYSIWYG) proposed so far in the literature. Second, we use optimal transport theory, and specifically the concept of the barycenter, to maximize decision maker utility under the chosen fairness constraints. Third, the algorithm is able to handle cases of intersectionality, i.e., of multi-dimensional discrimination of certain groups on grounds of several criteria. We discuss three main examples (credit applications; college admissions; insurance contracts) and map out the legal and policy implications of our approach. The explicit formalization of the trade-off between individual and group fairness allows this post-processing approach to be tailored to different situational contexts in which one or the other fairness criterion may take precedence. Finally, we evaluate our model experimentally.

Bias characterization, assessment, and mitigation in location-based recommender systems

“…In fact, RSs are known to be multi-stakeholder environments (Abdollahpouri et al 2019a), since they affect multiple actors in a direct way, mainly the users receiving the recommendations (consumers) and those behind the recommended objects (providers). Because of that, research on bias analysis and fairness measurements is needed; in particular, specific definitions, dependency variables, and mitigation approaches beyond those already studied for general Machine Learning (Zehlike et al 2020).…”

Fair Link Prediction with Overlapping Groups

“…The algorithm referred to as FLIP is a popular fair link prediction algorithm performing graph structural debias by reducing homophily present in the graph and measures fairness using reduction in modularity. Zehlike et al 40 . It is a optimal transport based score transformation algorithm to tackle intersectional attributes.…”

“…In this paper, we apply the score transformation approach to graph data by taking the former route and using quadratic programming to transform the scores to fair ones. We also use the algorithm of Zehlike 40 , that follows the other approach, as a baseline, by suitably adapting it to the graph data.…”

“…In the second approach, one can use the theory of optimal transport to find the distribution of scores corresponding to all the sensitive populations and optimally transport them to a fair score distribution. This approach has been followed in papers such as [36][37][38][39][40] . In this paper, we apply the score transformation approach to graph data by taking the former route and using quadratic programming to transform the scores to fair ones.…”

“…Originally, the theory began with the oppression of "black-women" in which two different forms of oppression were considered. In fair machine learning, while considering tabular data, intersectionality has received attention in classification 13,[42][43][44][45][46] , ranking 40 and subset selection 47 . Informally, in fair machine learning, we are in the intersectional setting if we have a set of sensitive populations that intersect among themselves e.g.…”