CKACL: A Collaborative Knowledge-Aware Contrast Learning Framework for Addressing Sparsity and Cold-Start Problems in Recommendation System

Knowledge graphs have emerged as a powerful tool for enhancing recommendation systems by encoding rich side information to model user preferences with improved accuracy. Despite their potential, most knowledge graph-based recommendation systems still face two persistent challenges: data sparsity and cold-start problems. To address these challenges, we propose a novel framework called Collaborative Knowledge-Aware recommendation system with Contrast Learning (CKACL), which effectively integrates collaborative information and knowledge-aware contrastive learning. Our approach derives user and item collaborative graphs, enriching the representation of the user preference with both direct interactions and global collaborative patterns. In addition, CKACL introduces an innovative structural consistency-guided contrastive learning mechanism, where perturbed knowledge graph views with high-order semantics are generated via consistency scores. Comprehensive experiments on three real-world datasets demonstrate CKACL’s superiority in addressing cold-start and data sparsity problems, outperforming state-of-the-art baselines with an average improvement of 20% in recall@20.