Modeling heterogeneous cross-domain interactions while ensuring reliable knowledge transfer remains challenging for cross-domain sequential recommendation. Existing methods often simplify cross-domain relations and overlook temporal evolution, which weakens the exploitation of fine-grained collaborative signals and yields rigid user representations. We propose DHT-MVC4Rec, a cross-domain sequential recommendation framework based on a dynamic heterogeneous graph and multi-view contrastive learning. The model builds a dynamic heterogeneous graph over cross-domain data and updates edge weights over time to capture long-term preference evolution. It further integrates sequential modeling with multi-view contrastive supervision to enhance cross-domain alignment while preserving domain-specific characteristics. Experiments on multiple cross-domain tasks show that DHT-MVC4Rec consistently outperforms strong baselines, improving NDCG@10 and HR@10 by 7.588% and 7.36% on average, respectively. Future work will explore dynamic heterogeneous graphs with different structural properties in diverse scenarios to further improve generalization. • Dynamic Heterogeneous Graph Construction: User preferences are modeled with time-evolving dynamic graphs that capture temporal behavior shifts and higher-order cross-domain interactions. • Preference-Level Multi-View Contrastive Learning: A multi-view contrastive learning framework is adopted to align user preference representations across domains, strengthening knowledge transfer and mitigating sparsity and noise. • Mechanism for updating dynamic weights of edges: A dynamic edge-weighting mechanism is introduced to adaptively reweight interactions over time, enabling both short-term interest modeling and long-term preference capture.
Wang et al. (Sun,) studied this question.