Learning Socially Optimal Information Systems from Egoistic Users.

Many information systems aim to present results that maximize the collective satisfaction of the user population. The product search of an online store, for example, needs to present an appropriately diverse set of products to best satisfy the different tastes and needs of its user population. To address this problem, we propose two algorithms that can exploit observable user actions (e.g. clicks) to learn how to compose diverse sets (and rankings) that optimize expected utility over a distribution of utility functions. A key challenge is that individual users evaluate and act according to their own utility function, but that the system aims to optimize collective satisfaction. We characterize the behavior of our algorithms by providing upper bounds on the social regret for a class of submodular utility functions in the coactive learning model. Furthermore, we empirically demonstrate the efficacy and robustness of the proposed algorithms for the problem of search result diversification.