A Graph Based Approach Towards Exploiting Reviews for Recommendation

Textual reviews, pervasive on many e-commerce websites, contain a lot of information. Many neural network models have been proposed to use the information of reviews to improve the performance of recommender systems. However, existing models usually use convolutional neural networks to learn the features of the reviews, often focus on the local interactions of words and lack the ability to capture long-distance and non-consecutive word interactions. Meanwhile, their ability should be strengthened on modelling the high-level interactions between users and items. Therefore, we propose a multi-view Graph based Approach towards exploiting Reviews for recommendation (GAR). It integrates the information of review content and user-item graph. In review view, we build an individual word co-occurrence graph for each review and use gated graph convolutional network to learn the features of reviews. In graph view, we use graph attention network to model high-order multi-aspect relations in the user-item graph. Both views use a graph based method. The representation of users and items learned from the two views are integrated to predict the final rating. Experiments on the benchmark datasets show that GAR achieves significantly better rating prediction accuracy compared to the state-of-the-art methods.