GLSCL: Graph local similarity contrastive learning for recommendation

Graph Contrastive Learning (GCL) has recently gained widespread adoption in recommendation systems owing to its outstanding performance and capability to alleviate data sparsity issues. GCL mitigates data sparsity issues by learning more uniformly distributed user and item representations. However, current recommendation approaches based on GCL have a significant drawback, which is to overlook the relationships between homogeneous nodes (i.e., users to users and items to items). The potential of contrastive learning in these contexts remains largely untapped because contrastive learning often focuses only on the user–item interaction space, missing the fine-grained, contextual similarities that exist within homogeneous nodes. These overlooked relationships can significantly improve recommendation accuracy by providing richer, more context-sensitive embeddings. To address this problem, we propose a Graph Local Similarity Contrastive Learning (GLSCL) framework, which enhances embedding uniformity and constructs contrast pairs based on local similarity. Specifically, to avoid the loss of original information, we employ a random perturbation contrastive task to enhance embedding uniformity and improve recommendation performance by exploring the inherent correlations among users (or items). GLSCL treats users (or items) with their local batch of most similar users (or items) as positive contrastive pairs during training, which can capture the homogenous relationships in user–user and item–item similarity relationships while maintaining embedding uniformity. To validate the proposed model, extensive experiments were conducted on three real-world datasets, including Douban-Book, Yelp, and Amazon-Book. On the three datasets, our model outperforms the suboptimal model with an average improvement of 3.23%, 3.28%, 3.55%, and 3.41% in four metrics, respectively. Extensive ablation experiments and visual analyses were conducted, providing conclusive evidence for the effectiveness of the proposed core modules.