SENGR: Sentiment-Enhanced Neural Graph Recommender

In recent years, reviews and user-item interaction have been recognized as valuable information to improve representation learning abilities in recommendations. However, on the one hand, the existing review-based recommendations normally ignore the importance of sentiment words regarding the corresponding aspect words, which reflect user preference for the item aspect. On the other hand, when modeling interaction, both user-user and user-item interactions should be considered. To solve these issues, in this paper, we propose a novel sentiment-enhanced neural graph recommender by incorporating the information derived from both textual reviews and bipartite graph. Specifically, we first design a hierarchically structured attention mechanism with a sentiment auxiliary task to help the recommendation task learn user preference for different aspects of items from reviews, where the co-attention mechanism is used to select important item/user reviews for the current user/item. Second, we construct a user-item interaction graph to capture preference-based user-item interaction with social-based user-user interaction, where the graph convolutional network is used to simulate the diffusion of information. Finally, we adopt a Factorization Machine model to accomplish the recommendation task. The experimental results demonstrate that our model significantly outperforms the related approaches w.r.t. rating prediction accuracy on Yelp and Amazon datasets.