Knowledge-Guided Semantically Consistent Contrastive Learning for sequential recommendation

Contrastive learning has gained dominance in sequential recommendation due to its ability to derive self-supervised signals for addressing data sparsity problems. However, caused by random augmentations (e.g., crop, mask, and reorder), existing methods may produce positive views with inconsistent semantics, which degrades performance. Although some efforts have been made by providing new operations (e.g., insert and substitute), challenges have not been well addressed due to information scarcity. Inspired by the massive semantic relationships in the Item Knowledge Graph (IKG), we propose a Knowledge-Guided Semantically consistent Contrastive Learning model for sequential recommendation (KGSCL). Specifically, we introduce two knowledge-guided augmentation operations, KG-substitute and KG-insert, to create semantically consistent and meaningful views. These operations add knowledge-related items from the neighbors in the IKG to augment the sequence, aligning real-world associations to retain original semantics. Meanwhile, we design a co-occurrence-based sampling strategy to complement knowledge-guided augmentations for selecting more correlated neighbors. Moreover, we introduce a view-target CL to model the correlation between semantically consistent views and target items since they exhibit similar user preferences. Experimental results on six widely used datasets demonstrate the effectiveness of our KGSCL in recommendation performance, robustness, and model convergence compared with 14 state-of-the-art competitors. Our code is available at: https://github.com/LFM-bot/KGSCL.