Multi-behavior recommendation systems leverage diverse user–item interactions (e.g., view, cart, purchase) to improve prediction accuracy, yet existing methods face critical limitations: they rely on pairwise graph structures that cannot capture high-order collaborative patterns, employ static edge weights that fail to adapt to heterogeneous interaction importance, and lack mechanisms to maintain consistency across behavioral contexts. We propose CDHMb (Contrastive Dynamic Heterogeneous Hypergraph for Multi-behavior Recommendation), a novel framework that addresses these challenges through three key innovations. First, we transform heterogeneous information networks into behavior-specific hypergraphs that explicitly encode semantic user communities based on shared item attributes (e.g., director, genre preferences). Second, we introduce a dual-level dynamic weighting mechanism that adapts both hypergraph structure (via attention-based hyperedge weights) and information propagation (via learned aggregation weights) to varying connection relevance. Third, we design a hierarchical contrastive learning objective that jointly optimizes local (within-behavior) discrimination and global (cross-behavior) coherence, ensuring embeddings capture both behavior-specific patterns and user-level identity. Extensive experiments on four real-world datasets (IMDB, Amazon, Yelp, Last.FM) demonstrate that CDHMb achieves substantial improvements over 14 state-of-the-art baselines, with gains of 7.5%–113% across Recall, Precision, and NDCG metrics. Comprehensive analyses validate CDHMb’s effectiveness in challenging scenarios including sparse data, cold-start users, and varying recommendation list lengths, while maintaining practical scalability for real-world deployment.
Alsaffar et al. (2026) studied this question.