Contrastive learning of similarity meta-path clustering for multi-behavior recommendation

Data sparsity has long hindered the performance of recommender systems. Leveraging multi-behavior data has emerged as a key strategy to alleviate this challenge. However, existing multi-behavior RS models often fall short in effectively capturing the structural and semantic representations of user and item behaviors, resulting in two unresolved issues: (1) how to construct semantically and structurally rich graphs that integrate multiple interaction types to mitigate data sparsity, and (2) how to uncover latent relationships among diverse behaviors, which are typically modeled independently. To address these challenges, we propose a novel model—Contrastive Learning based on Similar Meta-path Clustering (CSMC)—specifically designed for multi-behavior recommendation. CSMC performs contrastive learning (CL) from two complementary perspectives. First, we enhance the LightGCN framework to propagate both coarse-grained commonalities and fine-grained distinctions among user and item behaviors. Unlike traditional approaches that rely solely on user-item interaction pairs, our method treats multiple behaviors from the same user as positive samples and behaviors from different users as negatives, thereby improving the effectiveness of CL. Second, we introduce a similarity meta-path framework, which constructs a meta-path-based similarity graph through node-level similarity computation, allowing the model to transfer prior knowledge to low-resource tasks. Finally, CSMC jointly optimizes multi-behavior and meta-path contrastive objectives to extract both local and high-order semantic signals within a heterogeneous information network graph. Extensive experiments conducted on three real-world benchmark datasets—including ablation and sparsity analyses—demonstrate the superiority of CSMC, achieving average performance gains of 22.27% in Recall and 21.42% in NDCG compared to the strongest baselines.