This paper proposes CDGT, a framework addressing cold-start sequential recommendation by transferring knowledge from source domains at the group level. It identifies latent user groups with similar transition patterns across domains and establishes transferable group representations. The hierarchical architecture includes a Group Discovery Module, a Group-Aware Meta-Learner, and a Domain-Adaptation Layer. Experiments on real-world datasets show significant improvements over state-of-the-art baselines for cold-start users.
Key findings
CDGT transfers knowledge from source domains to target domains at the group level for cold-start sequential recommendation.
The framework identifies latent user groups with similar sequential patterns across domains.
A hierarchical architecture is proposed, including group discovery, group-aware meta-learning, and domain adaptation.
Extensive experiments demonstrate significant improvements over state-of-the-art baselines for cold-start users.
Limitations & open questions
The paper does not discuss the scalability of the CDGT framework for very large datasets.
The effectiveness of CDGT in diverse cultural or linguistic domains is not explored.