Journal of King Saud University Computer and Information Sciences General Computer Science

DG$$^{3}$$PR: dynamic global graph guided price-aware recommendation

Yingjie Kuang, Zhenhua Zeng, Ling Huang, Zhenyu Yang, Yuefang Gao, Chang-Dong Wang
Springer Nature · 2026-08-01 · DOI 10.1007/s44443-026-00469-0

Abstract

Abstract The price of items is an important factor for users’ purchase decisions. Current price-aware recommendation systems mainly focus on modeling user behavior and commodity attributes with the fixed price of items, often ignoring the impact of dynamic price fluctuation on user behavior. Additionally, these algorithms lack consideration of the temporal and global information of user-item actions, making it more difficult to capture user preferences for recent purchases. To address these limitations, we propose a novel Dynamic Global Graph Guided Price-aware Recommendation (DG 3 PR) method. Specifically, the proposed model comprises four main components: the Dynamic Subgraph Spatial Representation Learning (DSSRL) module, the Dynamic Subgraph Temporal Representation Learning (DSTRL) module, the Global Graph Spatial Representation Learning (GGSRL) module, and the prediction module. In the DSSRL module, a dynamic price grade discretization method is introduced to map item price to the corresponding grade. Subsequently, the DSTRL module captures the dynamic changes of commodities over time. Meanwhile, the GGSRL module learns the global information of the entire heterogeneous graph to capture the user’s purchase preferences globally. During prediction, the temporal and global contextual information is incorporated by the attention mechanism to learn the user’s purchase preferences. By performing such integration, the proposed model effectively facilitates the performance of price-sensitive recommendations in Business-to-Business (B2B) scenarios. To validate the proposed model’s effectiveness, we construct three real-world datasets from a B2B e-commerce platform and conduct extensive experiments. The results demonstrate that the proposed method outperforms the current competing methods by a substantial margin. Specifically, compared to the best-performing baseline LightGCN on the Middle region dataset, our model achieves improvements of 4.01%, 1.32%, and 3.92% in Precision@10, Recall@10, and NDCG@10, respectively, demonstrating its superior capability in capturing dynamic price-aware user preferences.

Read at publisher →