Meta-Learning based Heterogeneous Graph Attention Network for Top-N Review Recommendation

Abstract

User-generated content (UGC) has become more and more popular on the web and the published review is an essential type of UGC. Nevertheless, the explosion of reviews brings a problem of severe information overload. Therefore, most web services supply review recommendations for users. Traditionally, reviews of an item could be exhibited in chronological or popularity order without personalization. However, some researchers are aware of the significant role of the personalized review recommendation, which focuses on discovering users’ personalized preferences so that recommended reviews could match users’ preferences better. Unfortunately, it is hard to obtain users’ feedback on reviews due to the privacy protection and trade secrets. Furthermore, the difficulty in capturing varying patterns of users’ preferences and the sparsity of interactions between users and reviews are also challenging. To address these problems, we first formally define the top-N review recommendation problem and construct two categories of datasets based on a public dataset. Secondly, we propose a meta-learning based heterogeneous graph attention network incorporating multiple relationships among the users, items and reviews to model personalized users’ preferences and cope with the sparse situation. Moreover, to accelerate the message propagation computation, a method of the substructure-oriented local graph construction is proposed and is fused into the meta-learning framework based on a pair-wise ranking. For the top-N review recommendation, experiments are conducted on the two categories of a real-world dataset. Compared with the state-of-the-arts, the results validate the effectiveness of our model for review recommendation.