Online bipartite matching with partially bandit feedback

Inspired by the increasing popularity of online dating websites, we consider the problem of online bipartite matching. Previous work on this topic either assumes an unrealistic full-information feedback model or makes due with bandit feedback but requires us to solve a difficult computational problem on each online round. We take the middle-ground between these two extremes and propose a partially-bandit feedback model that is both realistic and allows us to design an efficient regret-minimizing algorithm. We present the ExpMatch online learning algorithm for bipartite matchings with partially-bandit feedback, analyze its computation complexity and its regret, and compare it to previous algorithms.