Bridging the semantic gap: LLM-enriched graph attention networks for robust cross-domain recommendation
APPLIED INTELLIGENCE, cilt.56, ss.1-17, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 56
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s10489-026-07585-4
- Dergi Adı: APPLIED INTELLIGENCE
- Derginin Tarandığı İndeksler: Applied Science & Technology Source, Information Science & Technology Abstracts (LISTA), Engineering Source (EBSCO), Scopus, Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest), Aerospace Database, Science Citation Index Expanded (SCI-EXPANDED), ABI/INFORM, Compendex, Educational research abstracts (ERA), INSPEC, Library, Information Science & Technology Abstracts (LISTA), zbMATH
- Sayfa Sayıları: ss.1-17
- Çukurova Üniversitesi Adresli: Evet
Özet
Cross-domain recommendation (CDR) aims to reduce sparsity and cold-start problems by transferring user preference signals across related domains. However, many graph-based CDR methods depend heavily on interaction structure or adversarial transfer, which can be unstable when the source and target domains differ semantically or when the target graph is sparse. This study proposes SG-GATv2-R, a semantic-guided residual graph attention framework for cross-domain recommendation. The model uses pretrained textual item representations as semantic anchors, constructs training-only user semantic profiles from observed interactions, and injects semantic compatibility into the GATv2 attention mechanism. It also introduces a learnable residual semantic–structural fusion layer to prevent graph propagation from overwriting useful semantic information. Cross-domain user alignment is modeled with an InfoNCE objective over shared users rather than adversarial adaptation. Experiments on Amazon cross-domain recommendation pairs show that SG-GATv2-R improves rating prediction and ranking performance over the selected baselines. Component-level ablations on Books → Electronics indicate that semantic initialization, semantic attention gating, graph propagation, and contrastive alignment contribute differently to performance, while the learned fusion parameters indicate that the model actively uses the residual semantic–structural pathway. Additional diagnostic analyses provide a cautious view of robustness. Under training-edge dropout, the model shows comparatively controlled degradation, although the NDCG@10 decrease at 30% dropout is not negligible. Shared-user overlap and metadata-availability sensitivity analyses further show that the model can tolerate reduced alignment and sparse metadata, but its performance still depends on metadata quality, graph connectivity, and available user overlap. These findings suggest that semantic-guided graph attention and residual semantic–structural fusion provide a useful but not assumption-free direction for cross-domain recommendation.