General-Special Neighborhood-Enhanced Graph Contrastive Learning for Recommendation

Abstract

Graph Neural Networks (GNNs) have gained widespread application in the field of recommendation systems due to their flexibility and effectiveness in modeling graph-structured data. GNN enable the collection of high-dimensional information from user-item interaction graphs through multiple layers of convolution and information propagation, facilitating a holistic modeling of users and items. However, GNN-based models face the challenge of over-smoothing due to sparsity, which hinders their ability to learn optimal node representations. To address this issue, graph contrastive learning algorithms have been introduced, yielding promising results. Yet, these graph contrastive learning-based methods tend to overlook the historical interactions.To tackle this problem, we propose a novel method called General and Specific neighborhood-enhanced Graph Contrastive Learning (GS-GCL). This method approaches node neighbor discovery and contrastive learning from both general and specific perspectives. We observe that in GNN even-pass information propagation encodes homogenous information that can be extracted as the general neighbor representation of the central nodes. Furthermore, we calculate the cosine similarity of a node's historical interactions to select the most similar nodes as its specific neighbors. To validate the effectiveness of our method, we conducted extensive experiments on three publicly available datasets. The results demonstrate significant improvements compared to the currently most effective method NCL method, particularly achieving a remarkable 14.14% enhancement on the Amazon-Books dataset. Our code repository can be found at: https://github.com/big-maomi/GS-GCL