Semantic-guided graph attention model improves cross-domain recommendation on Amazon data

Researchers have proposed a new framework for cross-domain recommendation that uses large language model–derived text representations to help recommender systems transfer user preferences between related product domains. The study, published in Applied Intelligence, introduces SG-GATv2-R, a semantic-guided residual graph attention framework designed to reduce sparsity and cold-start problems by moving preference signals across domains.

Cross-domain recommendation addresses a familiar weakness of recommender systems: when a target domain has few interactions, models struggle to suggest relevant items. Many existing graph-based approaches rely heavily on interaction structure or adversarial transfer, which the authors note can become unstable when the source and target domains differ semantically or when the target graph is sparse.

The new model takes a different route. It uses pretrained textual item representations as semantic anchors, builds training-only user semantic profiles from observed interactions, and injects semantic compatibility into the GATv2 attention mechanism. A learnable residual semantic–structural fusion layer is introduced to prevent graph propagation from overwriting useful semantic information during message passing.

Instead of adversarial adaptation, cross-domain user alignment is modeled with an InfoNCE contrastive objective computed over shared users. According to the abstract, this choice is intended to make the transfer of preference signals more stable when domains are only loosely aligned.

Experiments on Amazon cross-domain recommendation pairs show that SG-GATv2-R improves both rating prediction and ranking performance over the selected baselines. Component-level ablations on the Books to Electronics pair indicate that semantic initialization, semantic attention gating, graph propagation, and contrastive alignment each contribute differently to overall performance, and the learned fusion parameters suggest the model actively uses the residual semantic–structural pathway.

The authors also report diagnostic analyses that they describe as a cautious view of robustness. Under training-edge dropout, the model shows comparatively controlled degradation, although the decrease in NDCG@10 at 30 percent dropout is described as not negligible, meaning the ranking gains erode noticeably under that level of edge removal.

Further sensitivity analyses examine shared-user overlap and metadata availability. The model can tolerate reduced alignment between domains and sparse metadata, but its performance still depends on metadata quality, graph connectivity, and the amount of user overlap available between the source and target domains.

The study concludes that semantic-guided graph attention combined with residual semantic–structural fusion offers a useful but not assumption-free direction for cross-domain recommendation, with performance contingent on the quality of the underlying text representations and graph structure.

Subject of Research: A semantic-guided residual graph attention framework for cross-domain recommendation using LLM-enriched item representations

Article Title: Bridging the semantic gap: LLM-enriched graph attention networks for robust cross-domain recommendation

Article References: Özden, C., & Özcan, A. (2026). Bridging the semantic gap: LLM-enriched graph attention networks for robust cross-domain recommendation. Applied Intelligence, 56(16), Article 519. https://doi.org/10.1007/s10489-026-07585-4

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07585-4

Keywords: cross-domain recommendation, recommender systems, graph attention networks, large language models, GATv2, InfoNCE, contrastive learning, Amazon datasets, cold-start problem, semantic embeddings, NDCG, Applied Intelligence