CERA-OSR: Counterfactual Evidence Routing forAuditable One-Shot Subgraph Reasoning inSmart-City Knowledge Graph Completion

Knowledge graph completion supports data integration, while compact subgraph retrieval can overlook relevant evidence. CERA-OSR integrates train-only multi-view evidence pooling, relation-dependency priors, query-conditioned soft routing, uncertainty-triggered recovery and counterfactual evidence training within a one-shot subgraph reasoning framework. On WN18RR, FB15k-237, NELL-995 and YAGO3-10, CERA-OSR achieves MRR values of 0.626, 0.468, 0.603 and 0.676, respectively. The unweighted mean MRR is 0.5933, compared with 0.5738 for MoKGR, an absolute improvement of 0.0195. Ablation and graph-perturbation analyses examine component contributions and prediction sensitivity to selected evidence. Efficiency measurements reveal a trade-off between ranking performance and the cost of conditional recovery. The framework is motivated by smart-city knowledge graphs and evaluated on general knowledge graph benchmarks.