Neighbor importance-aware graph collaborative filtering for item recommendation

The emerging topic of Graph Neural Networks (GNN) has attracted increasing attention and achieved state-of-the-art (SOTA) performance in many recommendation problems, due to its strong ability in node representation with exploring high-order information. To learn a node's representation, previous meth-ods usually linearly combine the embeddings of node features, amusing the equal importance of neigh-bors. However, due to the intrinsic differences (i.e., degree, create time) over neighbors, we argue that these differences carry important signals for node representation. Ignoring them will lead to a suboptimal in node representation and thus weaken the effectiveness of the follow-up graph-based operations. To address it, we propose BIG-SAGE@ for item recommendation with rating prediction task, which is a neighbor importance-aware graph neural network. Specifically, its main idea is twofold: 1) A rating confidence-based neighborhood sampling method is introduced, making the sampling process biased to those more valuable nodes. 2) An attention network is integrated to achieve the rating prediction task, by flexibly incorporating information from user and item embedding features. Finally, we verified the effectiveness of the proposed model on six public data sets. Extensive experimental results demonstrate the superior performance of BIG-SAGE@ in the rating prediction and TopN ranking tasks, compared to the SOTA methods.& COPY; 2023 Elsevier B.V. All rights reserved.

Automated summary

The emerging topic of Graph Neural Networks (GNN) has achieved state-of-the-art performance in recommendation problems due to its strong ability in node representation. We propose BIG-SAGE@, a neighbor importance-aware GNN, for item recommendation and rating prediction. Through rating confidence-based neighborhood sampling and an attention network, BIG-SAGE@ outperforms SOTA methods in rating prediction and TopN ranking tasks.