Rating-oriented interest graph disentanglement for review-based recommendation

User rating prediction is a fundamental task in recommendation systems, whose performance critically depends on accurate modeling of fine-grained user preferences. Existing rating-aware methods, however, face two main limitations: they fail to adequately capture the interaction-dependent relationship between user rating behaviors and item rating distributions, and they treat numerical ratings as semantically uniform supervision signals, which introduces interference among heterogeneous preference intensities. To address these issues, we propose Rating-Oriented Interest Graph Disentanglement (RIGD), a graph convolutional framework that explicitly integrates rating semantics into preference modeling. RIGD introduces a statistically grounded rating deviation mechanism to characterize personalized preference boundaries derived from both user- and item-centric rating distributions. Based on these boundaries, the interaction graph is structurally disentangled into semantically coherent preference subgraphs-namely like, neutral, and dislike subgraphs-each modeling preference representations at distinct intensity levels and capturing rating-induced associations between users and items. Furthermore, a cross-subgraph fusion module is designed to jointly model and optimize local semantic preferences and global structural information. Extensive experiments on multiple benchmark datasets demonstrate that RIGD effectively mitigates feature interference, significantly improves rating prediction accuracy, and provides a principled approach for fine-grained preference modeling in recommendation systems.