RAWLSNET: Altering Bayesian Networks to Encode Rawlsian Fair Equality of Opportunity

Abstract

We present RAWLSNET, a system for altering Bayesian Network (BN) models to satisfy the Rawlsian principle of fair equality of opportunity (FEO). RAWLSNET's BN models generate aspirational data distributions: data generated to reflect an ideally fair, FEO-satisfying society. FEO states that everyone with the same talent and willingness to use it should have the same chance of achieving advantageous social positions (e.g., employment), regardless of their background circumstances (e.g., socioeconomic status). Satisfying FEO requires alterations to social structures such as school assignments. Our paper describes RAWLSNET, a method which takes as input a BN representation of an FEO application and alters the BN's parameters so as to satisfy FEO when possible, and minimize deviation from FEO otherwise. We also offer guidance for applying RAWLSNET, including on recognizing proper applications of FEO. We demonstrate the use of our system with publicly available data sets. RAWLSNET's altered BNs offer the novel capability of generating aspirational data for FEO-relevant tasks. Aspirational data are free from the biases of real-world data, and thus are useful for recognizing and detecting sources of unfairness in machine learning algorithms besides biased data.

Online Platforms and the Fair Exposure Problem under Homophily

“…Graph-based models of opportunity flow have considered similar, yet inherently different, fairness constraints and problems. For example, Liu et al (2021) consider fair equality of opportunity in settings where flow of opportunity proceeds along an acyclic graph and everyone is striving for the same desired outcome. Similarly, Arunachaleswaran et al (2021) approximately optimize social welfare in settings where opportunity flows along an acyclic graph.…”

The Equity Framework: Fairness Beyond Equalized Predictive Outcomes

“…For example, strategic classification (e.g, (35,36,37,38,39,40,41,42)) shows how people's reactions to the decision-making model highlights the differences in access through costs people incur. Although our framing of obstacles is quite analogous to budget framing in strategic classification, we formulate obstacles more generally, and the alleviation of obstacles strictly makes things better, that is to say, the obstacle-free feature values, z dominate the obstacle-refrained feature values x and the obstacle-free label y is not necessarily equal to the obstacle-refrained label, y. Relatedly, another body of work that highlights obstacles individuals face is causal and Bayesian inference (e.g., (16,17,18,19)). However, instead of ensuring equal access, the focus is mainly on redefining decision-making models, by, for example, changing accuracy metrics, weights of different features, or features used, among other interventions to qualify obstaclerefrained individuals.…”

“…Although several ML fairness research, for example, causal and Bayesian inference (e.g., (16,17,18,19)), fair decision-making (e.g., (20,21,22)) acknowledge obstacles individuals face, the focus is mainly on changing decision-making models, for example, to qualify obstacle-refrained individuals. We, however, argue that this only helps in the short and not the long term because unalleviated obstacles individuals face in accessing the model resurface in the model utilization, in which decision-makers evaluate individuals on how well they utilize the model as a form of feedback to the decision-makers.…”