Abstract
Graph Neural Networks (GNNs) have been successfully adopted in recommendersystems by virtue of the message-passing that implicitly captures collaborativeeffect. Nevertheless, most of the existing message-passing mechanisms forrecommendation are directly inherited from GNNs without scrutinizing whetherthe captured collaborative effect would benefit the prediction of userpreferences. In this paper, we first analyze how message-passing captures thecollaborative effect and propose a recommendation-oriented topological metric,Common Interacted Ratio (CIR), which measures the level of interaction betweena specific neighbor of a node with the rest of its neighbors. Afterdemonstrating the benefits of leveraging collaborations from neighbors withhigher CIR, we propose a recommendation-tailored GNN, Collaboration-Aware GraphConvolutional Network (CAGCN), that goes beyond 1-Weisfeiler-Lehman(1-WL) testin distinguishing non-bipartite-subgraph-isomorphic graphs. Experiments on sixbenchmark datasets show that the best CAGCN variant outperforms the mostrepresentative GNN-based recommendation model, LightGCN, by nearly 10\% inRecall@20 and also achieves around 80\% speedup. Our code is publicly availableat https://github.com/YuWVandy/CAGCN.