Anchor node guided global–local graph neural networks for multimedia recommendation

Abstract

Multimedia recommendation systems have garnered significant attention in recent years, particularly with GNN-based recommendation systems, which have demonstrated strong performance by aggregating information layer-by-layer through convolutional operations. However, previous studies have primarily focused on addressing the challenges of sparse interaction data, with limited attention given to issues arising from dense interactions. We argue that dense interactions hinder the accurate capture of user preferences, leading to bottleneck problem and an increased likelihood of noise. The introduction of GNNs further exacerbates these problems. Therefore, the problems arising from dense interaction cannot be ignored. In this paper, we propose a novel Anchor Node Guided Global–Local Graph Neural Networks for Multimedia Recommendation (AGG-LRec). To address the bottleneck problem caused by dense interactions, we propose a new graph neural network called anchor node module by using anchor nodes to aggregate the global information and interact with the target nodes. We also introduce two anchor-centric auxiliary objectives that explicitly regulate the semantic roles of anchor nodes. By enforcing anchor nodes to act as global semantic prototypes and maintaining cross-modal consistency at the anchor level, these objectives directly strengthen anchor-guided global aggregation under dense and noisy interactions. In addition, we propose a degree-sensitive edge pruning method for removing noise caused by dense interactions. This approach enhances the model’s robustness to noise by pruning edges. We also performed graph enhancement by introducing item–item edges. Experiments conducted on three datasets demonstrate that our proposed model outperforms baseline approaches. The source code is available at https://github.com/ren3570/AGG-LRec.