A Probabilistic Model for Diversifying Recommendation Lists

We propose a probabilistic method to diversify the results of collaborative filtering. Recommendation diversity is being studied by many researchers as a critical factor that significantly influences user satisfaction. Unlike conventional approaches to recommendation diversification, we theoretically derived a diversification method. Specifically, our method naturally diversifies a recommendation list by maximizing the probability that a user selects at most one item from the list. For enhanced practicality, we formulate a model for the proposed method on three policies — robust estimation, the use of only purchase history, and the elimination of any hyperparameters controlling the diversity. In this paper, we formally demonstrate that our method is practically superior to conventional diversification methods, and experimentally show that our method is competitive with conventional methods in terms of accuracy and diversity.