Influence-Guided Data Augmentation in Graph Contrastive Learning for Recommendation

Abstract

The graph contrastive learning approach, which combines graph convolutional networks (GCNs) with contrastive learning, has been widely applied in recommender systems and achieved tremendous success. Most graph contrastive learning (GCL) methods for recommendation perform random data augmentation operations on the user-item interaction graph to generate subgraphs, learn node embeddings through graph convolutional networks, and finally maximize the consistency of node embeddings in different subgraphs using contrastive loss. GCL improves the recommendation performance while slowing down the rate at which node embeddings tend to be similar, alleviating the over-smoothing problem to some extent. However, random data augmentation (e.g., random node dropout or edge dropout) will destroy the structure of the original input graph and change the original semantic information, leading to performance degradation. In this paper, we propose a novel graph contrastive learning model, IG-GCL, which uses the influence of elements in a graph to achieve guided data augmentation. Specifically, the model uses the mutual reinforcement network and node degree to calculate the importance scores of nodes and edges in the graph, respectively, thereby creating a more powerful data augmentation method to improve the performance of contrastive learning. We conduct extensive experiments on three real-world benchmark datasets. Experimental results demonstrate that IG-GCL can obtain performance improvements by stacking multi-layer neural networks, has the ability to mitigate the over-smoothing problem, and consistently outperforms the baseline, validating the effectiveness of the proposed influence-guided data augmentation method.