Reinforcement Learning to Diversify Top-N Recommendation

In this paper, we study how to recommend both accurate and diverse top-N recommendation, which is a typical instance of the maximum coverage problem. Traditional approaches are to treat the process of constructing the recommendation list as a problem of greedy sequential items selection, which are inevitably sub-optimal. In this paper, we propose a reinforcement learning and neural networks based framework – Diversify top-N Recommendation with Fast Monte Carlo Tree Search (Div-FMCTS) – to optimize the diverse top-N recommendations in a global view. The learning of Div-FMCTS consists of two stages: (1) searching for better recommendation with MCTS; (2) generalizing those plans with the policy and value neural networks. Due to the difficulty of searching over extremely large item permutations, we propose two approaches to speeding up the training process. The first approach is pruning the branches of the search tree by the structure information of the optimal recommendations. The second approach is searching over a randomly chosen small subset of items to quickly harvest the fruits of searching in the generalization with neural networks. Its effectiveness has been proved both empirically and theoretically. Extensive experiments on four benchmark datasets have demonstrated the superiority of Div-FMCTS over state-of-the-art methods.