Deep Learning for Two-Sided Matching Markets

Abstract

In two-sided matching, we seek to create matchings between agents on two sides of a market, each of whom has ranked, ordinal preferences over agents on the other side. Of particular interest are stable matchings, i.e., matching for which no pair of agents would mutually prefer to be matched to each other than to their assigned partners. The first part of this thesis gives an introduction to the theory of two-sided matching, focusing on methods of finding stable matchings, the structure of the set of stable matchings, and questions of strategic behavior under matching mechanisms. The second part is a study of the use of deep learning for questions that fundamentally lie within the realm of economic theory. I introduce a deep learning framework that can be used to model one-to-one matching markets, and demonstrate its ability to learn approximately stable matching mechanisms, replicating the original stability results as known by theoreticians, while also finding complete matches and achieving high aggregate welfare. The framework holds promise for examining the trade-offs between dominant-strategy incentive compatibility, a strategic question for individual agents, and the stability of the overall matching.