RI-GCN: Review-aware Interactive Graph Convolutional Network for Review-based Item Recommendation

Abstract

A wealth of semantic features exist in the reviews written by users, such as rich information on item features and implicit preferences of users. Existing review-based recommendation models usually employ Convolutional Neural Networks (CNNs) to learn representations of users and items from reviews. However, these CNNs-based models suffer from two main problems: (1) they only consider the information of the word itself during the convolution, ignoring the high-order contextual semantic information of the word; (2) they model user/item attributes in a static and independent way, ignoring the potential feature interaction between them. Therefore, we propose a novel Review-aware Interactive Graph Convolutional Network (RI-GCN) for review-based item recommendation. Specifically, we design a Review-aware GCN component to model the message propagation of graphs constructed from reviews, capturing the contextual features of words. A feature interactive GCN component is then proposed to capture the user/item high-order collaborative features in the user-item graph, enabling the model to further complement and refine u ser/item a ttributes. Finally, we adopt a Factorization Machine model for the recommendation task. Experimental results demonstrate that the proposed model is superior to state-of-the-art models.