C-GDN: core features activated graph dual-attention network for personalized recommendation

As a popular graph learning technique, graph neural networks (GNN) show great advantages in the field of personalized recommendation. Existing GNN-based recommendation methods organized user-item interactions (e.g., click, purchase, review, etc.) as the bipartite graph and captured the higher-order collaborative signal with the aid of the GNN to achieve personalized recommendation. However, there exists two limitations in existing studies. First, core features activating user-item interactions were not be identified, which causes that user-item interactions fail to be accurately exploited at the feature level. Second, existing GNNs ignored the mutual association among neighbors in information propagation, which results in structural signal in the bipartite graph not being sufficiently captured. Towards this end, we developed the core features activated graph dual-attention network, namely C-GDN, for personalized recommendation. Specifically, C-GDN firstly identifies core user and item features activating user-item interactions and employs these core features to initialize the bipartite graph, which effectively optimizes the utilizing of user-item interactions at the feature level. Furthermore, C-GDN designs a novel graph dual-attention network to conduct information propagation, which captures more sufficient structural signal in the bipartite graph by considering information from neighbors as well as their mutual association. Extensive experiments on three benchmark datasets shows that C-GDN outperforms state-of-the-art baselines.