Research map / Graph learning / Knowledge and temporal graphs
DuetGraph: Coarse-to-Fine Knowledge Graph Reasoning with Dual-Pathway Global-Local Fusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DuetGraph, a KG reasoning model that separates local message passing and global attention into two parallel pathways fused adaptively, plus a coarse-to-fine top-k reranking to counter score over-smoothing.
Problem
Knowledge graph reasoning methods that stack message passing and attention in one stage suffer from score over-smoothing, where incorrect entities receive scores close to correct ones, limiting discrimination and reasoning quality.
Approach
Decouple local (message passing) and global (attention) processing into two parallel pathways fused by a learnable weight alpha, and apply a coarse-to-fine scheme that splits the entity score table into high- and low-score subsets and selects the final answer by a threshold on the gap between their top scores.
Key results
- On 12 inductive subsets, DuetGraph improves MRR and Hits@1 over baselines, e.g., up to 8.6% MRR and 8.7% Hits@1 gains, ranking first in Hits@1 on all FB15k-237, WN18RR, NELL-995 versions.
- On four transductive datasets, DuetGraph reaches MRR 0.453 on FB15k-237 and 0.593 on WN18RR, outperforming KnowFormer (0.430 and 0.579).
- Ablation shows removing the dual-pathway fusion model causes the largest drop (FB15k-237 MRR 0.453 to 0.355), and the method reaches 1.8x training acceleration claimed on FB15k-237 and YAGO3-10.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the local pathway (ablation) | DuetGraph · FB15k-237, WN18RR, NELL-995, and YAGO3-10 | MRR: 0.445 -> 0.453; 0.584 -> 0.593; 0.582 -> 0.590; 0.617 -> 0.631 | Helped |
| add the global pathway (ablation) | DuetGraph · FB15k-237, WN18RR, NELL-995, and YAGO3-10 | MRR: 0.441 -> 0.453; 0.565 -> 0.593; 0.586 -> 0.590; 0.614 -> 0.631 | Helped |
| add coarse-to-fine reasoning (ablation) | DuetGraph · FB15k-237, WN18RR, NELL-995, and YAGO3-10 | MRR: 0.437 -> 0.453; 0.580 -> 0.593; 0.567 -> 0.590; 0.616 -> 0.631 | Helped |
| add dual-pathway fusion (ablation) | DuetGraph · FB15k-237, WN18RR, NELL-995, and YAGO3-10 | MRR: 0.355 -> 0.453; 0.512 -> 0.593; 0.534 -> 0.590; 0.563 -> 0.631 | Helped |
| add threshold Δ to correct low-score predictions (ablation) | DuetGraph · FB15k-237, WN18RR, NELL-995, and YAGO3-10 | MRR: 0.395 -> 0.453; 0.551 -> 0.593; 0.544 -> 0.590; 0.595 -> 0.631 | Helped |
| use a triplet-based coarse model rather than message passing or transformer coarse models (ablation) | DuetGraph · FB15k-237, WN18RR, NELL-995, and YAGO3-10 | MRR: 0.446 -> 0.453; 0.589 -> 0.593; 0.584 -> 0.590; 0.622 -> 0.631 | Helped |
| vary the high-score subset size k (ablation) | DuetGraph · Four transductive datasets | performance metrics | No clear effect |
| use DuetGraph instead of KnowFormer | Transductive KG reasoning · FB15k-237, WN18RR, NELL-995, and YAGO3-10 | MRR: 0.430 -> 0.453; 0.579 -> 0.593; 0.566 -> 0.590; 0.615 -> 0.631 | Helped |
Surprising findings
- Language-model-based SimKGC and MoCoKGC score very high on WN18RR, and the authors estimate over 70% of WN18RR entity texts may appear in the BERT pretraining data, suggesting possible data leakage.
- Using the triplet-based HousE as the coarse model works best, which the authors attribute to maximal architectural diversity between coarse and fine models.
Limitations
- Theoretical bounds rely on simplifying assumptions (e.g., spectral bounds on weight matrices and a Gaussian-like score distribution), and the empirical score-gap claims are not fully validated.
- Coarse-to-fine selection depends on a threshold Delta and top-k cutoff, requiring tuning.
- Several baselines are missing on some datasets because of code or data limitations, so comparisons are incomplete.
Experimental setup
Datasets: FB15k-237 WN18RR NELL-995 YAGO3-10 UMLS FB13 WN11 Wikidata5M Freebase
Benchmarks: FB15k-237 WN18RR NELL-995 YAGO3-10 UMLS FB13 WN11 Wikidata5M Freebase
Metrics: MRR Hits@1 Hits@10 accuracy
Baselines: DRUM NBFNet RED-GNN A*Net AdaProp Ingram KnowFormer SimKGC HousE HittER SAttLE MetaSD
Builds on: KnowFormer NBFNet HousE RED-GNN AdaProp
Closest papers by meaning
- Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models (ICML 2025 · similarity 0.88)
- Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning (ICLR 2024 · similarity 0.88)
- Towards Graph Foundation Models: Training on Knowledge Graphs Enables Transferability to General Graphs (NeurIPS 2025 · similarity 0.88)
- Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties (NeurIPS 2025 · similarity 0.88)
- Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph (ICLR 2024 · similarity 0.87)
- Plan-on-Graph: Self-Correcting Adaptive Planning of Large Language Model on Knowledge Graphs (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.