Learning to Recommend via Inverse Optimal Matching

We consider recommendation in the context of optimal matching, i.e., we need to pair or match a user with an item in an optimal way. The framework is particularly relevant when the supply of an individual item is limited and it can only satisfy a small number of users even though it may be preferred by many. We leverage the methodology of optimal transport of discrete distributions and formulate an inverse optimal transport problem in order to learn the cost which gives rise to the observed matching. It leads to a non-convex optimization problem which is solved by alternating optimization. A key novel aspect of our formulation is the incorporation of marginal relaxation via regularized Wasserstein distance, significantly improving the robustness of the method in the face of observed empirical matchings. Our model has wide applicability including labor market, online dating, college application recommendation. We back up our claims with experiments on both synthetic data and real world datasets.