IHG4MR: Interest-oriented heterogeneous graph for multirelational recommendation

Abstract

Multirelational recommendation usually considers multiple types of user–item interaction data. Recently, graph convolutional networks (GCNs) have been widely used for recommendation systems. However, most existing GCN-based recommendation models only utilize one type of interactive data, ignoring other types of interactive data. At the same time, it is worth noting that GCN-based recommendation models are prone to oversmoothing problems. This means that when GCN-based recommendation models stack more layers, they also import noisy messages from users who have no common interest with the target user, resulting in a recommendation performance bottleneck. Consequently, it is necessary to reform the traditional GCN to make it suitable for a multirelational recommendation. A novel multirelational recommendation model through an interest-oriented heterogeneous graph (IHG4MR) is proposed in this paper. To effectively capture information from high-level nodes in a heterogeneous graph, IHG4MR first utilizes user characteristics and graph structure to split users who have similar preferences and then employs these segmented users and their interacting items to build multiple new interactive graphs instead of the previous whole graph to implement graph convolution operations. IHG4MR reduces the negative information carried during embedding propagation of the graph convolutional layers, thereby mitigating the effect of oversmoothing. Extensive experimental results on two real world datasets show that IHG4MR generally outperforms existing state-of-the-art models, demonstrating the effectiveness of our model.