Dual channel group-aware graph convolutional networks for collaborative filtering

Jinsong Zhao, Kaiwen Huang, Ping Li

Abstract: Among the existing collaborative filtering approaches, graph-based collaborative filtering models that adopt graph convolutional networks for representation learning have achieved state-of-the-art results. However, most graph convolutions on the user-item graph treat all neighbors equally, while different users may have different interests and different items can belong to different categories, aggregating information about all neighbors without discrimination can lead to performance degradation. In addition, the user-item interaction graph is essentially a heterogeneous graph, but most of the existing models regard it as homogeneous, which also can lead to sub-optimal results. To address the above issues, we design a dual channel group-aware graph convolution model, called DG-GCN. Specifically, it first performs message passing on the user-item interaction graph to leverage the direct and higher-order connectivity information for further grouping, and then groups users and items separately through dual group-aware modules based on their latent interests and categories. This scheme allows nodes to learn from more similar homogeneous nodes, thus preventing noisy information from participating in the propagation process. Experimental results show that DG-GCN outperforms the state-of-the-art GCN model on all five benchmark datasets, with up to 5\(\%\) and 4\(\%\) relative improvements over LightGCN on Recall@20 and NDCG@20, respectively.