Multi-Modal Relational Side Information Graph Attention Networks for Recommender System

With the rapid growth of online information, recommender systems have become an effective strategy for handling information overload. One of the most important factors affecting the user's purchase activities is the side information of the item. Recently, some research has shown that analyzing the side information (e.g., visual and textual information) in multiple modalities can improve the performance of the recommender system. Existing works on multimedia recommendation using graph neural networks largely depend on the interaction records, while less effort focuses on the deeper relationships between interactions and different types of side information. In this paper, we propose a novel graph structure by combining the edges of user-item interactions and relational edges of side information in different modalities. Each modality contains a specific graph structure to learn the representations of users and items to make the recommendation. We design a Multi-modal Relational Side Information Graph Attention Networks (MRSI-GAT) framework built upon the message-passing with attention mechanism of graph attention networks. We conduct experiments on two public datasets, Amazon and MovieLens, and the results of the model outperform the state-of-the-art methods.