Relation-aware graph attention network for intelligent retrieval over compact domain knowledge graphs

Abstract

How well can graph-neural retrieval work when the knowledge graph has only a few hundred entities? Specialised engineering domains—aquaculture processing, pharmaceutical formulation, precision manufacturing—produce compact yet semantically dense graphs that starve collaborative filtering and defeat keyword rankers blind to multi-hop relational chains. R-KGAT tackles this question with a three-stage pipeline: LLM-assisted KG construction (∼200 entities, ∼500 typed relations), TransE/DistMult embedding, and a relation-aware two-layer GAT whose attention coefficients sit on named relations, making every retrieval result interpretable by construction. On an aquaculture-domain testbed (50 users, 180 items) the pipeline reaches Hit@10 = 0.782 and MRR= 0.463, 8–25% above all baselines. Ablation reveals a split that surprised us: graph structure alone accounts for 13.7 points; GAT message-passing adds 8.4. No post-hoc interpretation (LIME, SHAP) is needed.