Enhancing Knowledge-Aware Recommender Systems with Multi-Level Cross-View Contrastive Learning

Knowledge graphs (KGs) are becoming increasingly essential in recommender systems, with Graph Neural Network (GNN)-based models gaining substantial attention in knowledge-aware recommendation (KGR). Among these, the Collaborative Knowledge-Aware Attentive Network (CKAN) has shown impressive performance by using a heterogeneous propagation strategy to explicitly integrate collaborative filtering (CF) and KG data, alongside a knowledge-aware attention mechanism to differentiate the contributions of various knowledge neighbors. However, CKAN faces two primary challenges: (1) limited capacity to capture complex feature interactions and (2) imbalanced information usage due to sparse user-item interactions in CF and redundant knowledge within the KG component. To address these challenges, we propose an enhanced version of CKAN, named MCEKR. Our approach introduces a novel multi-level cross-view contrastive learning (MCCLK) technique that strengthens representation learning from diverse perspectives and extracts latent semantic information from the KG. This strategy not only improves CKAN’s representation power and training efficiency but also reduces reliance on large-scale datasets. Furthermore, we design a knowledge-aware fine-grained attention mechanism to capture higher-order collaborative signals and refine user preference modeling. This mechanism dynamically adjusts attention weights for entities and relational paths, facilitating better semantic disambiguation. Extensive experiments conducted on four public datasets demonstrate that MCEKR significantly outperforms state-of-the-art methods, especially in scenarios with sparse user-item interactions.