Knowledge&Social-based collaborative method with contrastive graph structure learning for explainable recommendation

Abstract

Explainable Recommendation has attracted increasing attention due to the growing significance of data privacy and model security in recommendation systems. However, the effectiveness of robust and security-sensitive recommendation methods may be constrained by limited observed data, potentially resulting in suboptimal accuracy and reliability. Although introducing multi-source side information helps mitigate data sparsity issues and improve recommendation performance, it also presents new challenges, including semantic disparities and noise interference. In view of these observations, we propose a Knowledge&Social-based collaborative method with Contrastive Graph Structure Learning for explainable recommendation, named KSCGSL. It establishes multi-view representations for users and items with explainable learning based on knowledge-enhanced semantic-aware modeling and social network-driven preference learning, both refined via contrastive graph structure optimization. Specifically, KSCGSL introduces a dual graph augmentation mechanism based on knowledge graph and semantic awareness for item embedding learning. For user modeling, it captures user preferences from user-item interaction analysis and augments them through social relations. To solve the inherent semantic inconsistencies across multiple views and mitigate noise interference, contrastive graph structural learning is incorporated to optimize embedding learning and filter structural noise. Experiments conducted on three publicly available datasets demonstrate that KSCGSL achieves significant improvements in recommendation accuracy with explainable manners.