Heterogeneous Graph Transfer Learning for Diversity-enhanced Cross-Domain Sequential Recommendation

Cross-domain Sequential Recommendation (CDSR) aims to alleviate the data sparsity problem in recommender systems by leveraging auxiliary information from other domains to better capture users’ sequential preferences. However, existing CDSR methods still face several limitations. Most approaches rely heavily on overlapping users to facilitate knowledge transfer, which constrains their ability to fully explore cross-domain item transition patterns. Moreover, they struggle to generalize to multi-domain scenarios and primarily optimize for recommendation accuracy. To address these issues, this paper introduces item category attributes into CDSR, defines the problem of Category-aware CDSR (CCDSR), and proposes a Heterogeneous Graph Transfer Learning model (HGTL) along with its diversity-enhanced variant HGTL-D. Specifically, we construct a cross-domain heterogeneous graph that connects user, item, and category nodes across different domains, where edges are expanded based on semantic similarity and category co-occurrence to enhance the flexibility of message propagation. In addition, we design meta-paths from multiple perspectives—item, user, and category—to guide information aggregation, enabling effective knowledge transfer across domains while reducing dependence on overlapping users. Furthermore, we model users’ item-level sequential preferences in each domain through attention modules and introduce an item-category attention layer to capture category-level sequential preferences. Then, we propose a personalized gating mechanism to adaptively transfer knowledge between domains and predict users’ favorite items in each domain. Finally, we develop a personalized MMR-based re-ranking module that quantifies users’ diversity tendencies via the category entropy of their historical interactions and dynamically adjusts the trade-off between accuracy and diversity in the final recommendations. Extensive empirical studies on three real-world datasets show that our HGTL achieves significant improvements in accuracy compared with the state-of-the-art baselines and supports multi-domain scenario expansion. Meanwhile, our HGTL-D further enhances diversity while retaining accuracy in cross-domain sequential recommendation.