Answers are grounded in this paper and cite their sources.
Personalized recommender systems are essential for mitigating information overload. However, existing approaches face challenges in fine-grained preference modeling, behaviorally relevant explanations, and cross-domain transfer. We propose Aspect-Grounded Heterogeneous Graph Learning (AGHGL), a recommendation framework that uses aspect-level sentiment semantics as a common basis for preference modeling, explanation, and cross-domain transfer. AGHGL constructs a heterogeneous sentiment graph integrating users, items, aspects, and opinion expressions, enabling relation-aware message passing and aspect-mediated preference modeling. A margin-based intervention-sensitivity objective couples attention-derived aspect importance with prediction changes under aspect masking, aligning explanations with prediction behavior. A dual-level aspect alignment mechanism facilitates cross-domain transfer by aligning aspect representations at semantic and graph-structural levels. Experiments on real-world datasets show that AGHGL consistently outperforms the evaluated baselines in recommendation accuracy, with larger relative gains under sparse and cold-start conditions in the target domain. Human evaluation indicates improved perceived explanatory quality, while intervention-based analysis shows that higher aspect importance is associated with larger prediction changes. These results suggest that the aspect-grounded framework provides a useful basis for jointly modeling recommendation, explanation, and cross-domain transfer within the evaluated settings.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: