The Price of Fairness in Bipartite Matching

We investigate notions of group fairness in bipartite matching markets involving agents and jobs, where agents are grouped based on sensitive attributes. Employing a geometric approach, we characterize how many agents can be matched in each group, showing that the set of feasible matchings forms a (discrete) polymatroid. We show how we can define weakly-fair matchings geometrically, for which poly-matroid properties imply that they are maximal. Next, we focus on strong fairness notions (inspired by group-fairness metrics in machine learning), where each group gets their exact same fraction of their entitlement, and we explore the Price of Fairness (PoF), i.e., the loss in optimality when imposing such fairness constraints. Importantly, we advocate for the notion of opportunity fairness, where a group entitlement is the maximum number of agents that can be matched without the presence of other competing groups. We show that the opportunity PoF is bounded independently of the number of agents and jobs, but may be linear in the number of groups. Finally, we provide improved bounds with additional structural properties, or with stochastic graphs.