Mutual Attention Graph Neural Network Based on Joint Representation of Nodes and Reviews for Recommendation

Recently, recommendation system have achieved good results by applying graph neural network to user-item inter-action graph. However, current graph neural network mainly deals with structured data and cannot deal with unstructured re-view text well. Item reviews are a unique way for users to choose to purchase the item. Therefore, combining a user-item interaction graph with related review text will obtain better recommendation performance. At the same time, most of the recommendation methods that have been proposed simply concatenate the representations from different modalities to make predictions. This can-not take advantage of the information from different modalities. To solve these problems, we propose a Mutual Attention graph neural Network (MAN) for personalized recommendation. MAN first extracts user/item node representation on user-item interaction graph through node feature extraction module, and extracts user/item review text representation through review feature ex-traction module. Then a mutual attention module is used to correlate node representation and review text representation, so as to capture the correlation between the node representation and the review text representation during the training process. Experimental results on three real-world datasets show MAN is better than the state-of-the-art personalized recommendation method.