REDRL: A review-enhanced Deep Reinforcement Learning model for interactive recommendation

Abstract

Recent advances in interactive recommender systems (IRS) have received wide attention due to its flexible recommendation strategy and optimization for users’ long-term utility. Considering this interaction paradigm of IRS, researchers have made some attempts to incorporate reinforcement learning (RL) models into IRS, because of the excellent ability of RL in long-term optimizing and decision making. However, data sparsity is an intractable problem most IRS urgently need to address. Although a small amount of work has exploited reviews to address data sparsity, they ignored the varying importance of items for modeling the user. In addition, most existing RL-based approaches suffer from decision-making difficulties when the action space becomes large. To solve above problems, in this work, we present a Review-enhanced Deep Reinforcement Learning model (REDRL) for interactive recommendation. Specifically, we utilize text reviews, combined with a pretrained review representation model to acquire item review-enhanced embedding representations. Then we formalize the recommendation problem as a Markov Decision Process (MDP), and exploit deep reinforcement learning (DRL) to model the interactive recommendation. Notably, we introduce a multi-head self-attention technique to capture distinct importance of various items in the sequence behavior, which is overlooked by existing work when modeling the user preference. In this way, we can model long-term dynamic preferences of users accurately and discriminately for comprehensive interactive recommendation. Moreover, we subtly combine the semantic structure information in the user–item bipartite graph with meta-paths in heterogeneous information networks (HIN), to filter some irrelevant items and obtain candidate items dynamically. By this means, the size of the discrete action space is effectively reduced from a new anger. The experimental results based on three benchmark datasets demonstrate the efficiency of our method with significant improvement over state-of-the-art.