Hypergraph-structured variational inference for next-basket recommendation
Author Identifier (ORCID)
Abstract
Next-basket recommendation (NBR) aims to suggest a basket of items for the user's next purchase. The key issue in NBR is not only modeling the complicated relationships between baskets and items but also capturing the diverse purchasing intents within baskets. Existing methods extensively employ hypergraphs for the effective modeling of basket-item relationships in NBR. However, they rely on fixed-point basket/item vectors, which may suffer from a capacity bottleneck in preserving the diversity of intent semantics during propagation. Instead, a more rational approach employs variational representations to model baskets and items, while learning their associative relations through hypergraph structures. Nevertheless, applying variational learning to hypergraphs is challenging due to inter-distribution transmission interference and inference noise. To tackle these challenges, this paper proposes a HyperGraph-structured Variational model, named VarHG, to learn diverse purchasing intents by capturing item-basket relations. Specifically, we carefully design a hypergraph-driven basket multi-intent learning module that integrates a hypergraph-structured variational network with adaptive link-weight refinement to mitigate inter-distribution propagation interference, and a hyperedge-driven variational encoder to infer baskets' multi-intent distributions from hyperedge features. Moreover, we propose a multi-intent-aware sequential preference learning module that models users' diverse interests from basket sequences in a distributional manner. Besides, we introduce a learned accept-sampling mechanism that employs a fake-basket discriminator to filter out noisy samples, ensuring reliable intra-basket item sampling for next-basket prediction. To enhance efficiency, we propose VarHG-E, an efficient version of VarHG built upon anchor-based hypergraph techniques. Finally, we conduct experiments on real-world datasets to demonstrate VarHG's superior performance compared with other baselines.
Keywords
hypergraph learning, next-basket recommendation, variational inference
Document Type
Journal Article
Date of Publication
1-1-2026
E-ISSN
15582191
ISSN
10414347
Publication Title
IEEE Transactions on Knowledge and Data Engineering
Publisher
IEEE
School
School of Business and Law
Copyright
subscription content
Recommended Citation
Yang, W., Liu, Z., Zhou, R., Chen, L., Li, J., Liu, C., & Xu, J. (2026). Hypergraph-structured variational inference for next-basket recommendation. IEEE Transactions on Knowledge and Data Engineering. Advance online publication. https://doi.org/10.1109/TKDE.2026.3720221