CRII: AF: RUI: Algorithmic Fairness for Computational Social Choice Models

Wellesley College, Wellesley Hills MA

Investigators

Abstract

Computers now support many kinds of preference aggregation/voting systems and have even made new ones possible, driving deeper study into computational social choice. For example, online liquid democracy platforms allow users to vote directly on a topic or delegate their vote to a trusted proxy whom they believe is more informed and will represent their interests. These systems have allowed businesses to make collective decisions on issues ranging from product design to food offered in the corporate cafeteria. However, recent audits of algorithms used in machine learning and artificial intelligence systems have taught us that algorithmic decisions made by computers have the potential to unintentionally disadvantage individuals or groups of people. Such algorithmic bias and discrimination can be countered by designing algorithms with specific fairness guarantees built in. This project will investigate computational social choice through an algorithmic fairness lens and illuminate the theoretical limitations of achieving difficult or conflicting concepts of fairness. In addition to advancing knowledge that benefits society and the research community, topics studied in this project will be used to enrich courses at every level of the computer science curriculum with engaging real-world applications, and lessons featuring these topics will be shared with the broader computer science education community. Finally, the investigator will mentor undergraduate students from traditionally underrepresented groups in computer science, diversifying the pipeline to graduate school. Algorithms are commonly used to implement existing and proposed preference aggregation/voting systems as well as to analyze them. At the same time, algorithmic bias and discrimination has been documented in a broad range of applications from hiring to medicine to criminal justice. In many of these areas, the research community has responded by formalizing computational definitions of fairness and designing algorithms that explicitly offer fairness guarantees, especially for machine learning tasks such as classification or recommendation. At a high level, this project seeks to unite computational social choice and the recent research into algorithmic fairness and fairness, accountability, and transparency (FAccT) in automated systems more broadly. The main contributions to computer science and other disciplines will be: (1) Formulating new computational problems, objectives, and constraints for implementing and evaluating voting systems that can guide future work in algorithmic fairness that is grounded in a specific real-world application; (2) Designing and analyzing algorithms for these problems that can provide fairness guarantees; and (3) Proving impossibility results that establish which notions of fairness in these settings are incompatible with each other or intractable. A focus of (1) will be to build connections between foundational, theoretical work in algorithmic fairness and specific application areas in computational social choice. The work of (3) will echo the seminal impossibility results in the areas of both algorithmic fairness and social choice theory. Thus, (1) and (3) will inform the investigator’s own work on (2), but also pose new problems to the research community. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.