HSP-GCN: Historical Semantic Proxy Graph Convolutional Networks for Review-Aware Recommendation
Year: 2026
Type: conference-paper
Abstract: In rating prediction tasks, review text provides an indispensable semantic dimension for capturing users’ complex preferences. However, in the pursuit of higher predictive accuracy, existing studies often overlook a stringent yet practically critical constraint: at the real-world inference stage, the review associated with the target edge has not yet been generated and is therefore completely unavailable. This inference-time constraint leads to substantial performance degradation for many models that rely on target-specific semantic information. To address this issue, we propose HSP-GCN (Historical Semantic Proxy Graph Convolutional Network). Instead of attempting to exploit non-existent future information, HSP-GCN reconstructs semantic associations by leveraging the full set of non-target historical reviews accumulated by users and items. During our investigation, we observe that, compared with carefully designed target-specific retrieval mechanisms, proxy representations derived from global historical mean aggregation exhibit superior stability and broader semantic coverage. Extensive experiments conducted on four Amazon benchmark datasets demonstrate that, without relying on target-edge reviews, HSP-GCN achieves significant performance gains over strong baseline models.
Authors: Junxian Wu, Zhaoqian Zhong
Institution: Dalian University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204840586