Research map / Graph learning / Knowledge and temporal graphs
Rule-Guided Graph Neural Networks for Explainable Knowledge Graph Reasoning
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 a monotonic GNN for KG reasoning whose first- and second-layer parameters encode Horn rules, so it can be trained with rule-confidence guidance and rules can be read off its parameters.
Problem
Rules extracted from explainable GNNs for knowledge graph reasoning have low standard confidence, so their plausibility is not verified, and existing rule-injection methods do not give faithful symbolic explanations.
Approach
A monotonic GNN with non-negative parameter matrices over pair and singleton nodes, where specific rule forms (R1-R8) correspond to parameter thresholds; the model is regularised toward rules' standard confidences and rules are extracted by reading off parameters, with soundness proven.
Key results
- On inductive link prediction over 9 benchmarks, the proposed model attains the best AUC on most datasets (e.g., 99.6 on NELL-995 v1 and INDIGO-BM) and highest F1 on most, while remaining competitive in precision.
- Over 90% of extracted rules are high-quality (standard confidence >= 0.4) in most cases, higher than NCRL, MMDRUM, MGNN and INDIGO.
- Most injected high-quality rules are recovered after training, with the model also discovering additional high-quality rules; applying all rules together reduced F1 on FB15K-237 (83.4 vs 90.1 for v1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add rule guidance (ablation) | Rule-Guided GNN · FB15K-237 v1–v4 | F1 score: 88.6 -> 90.1; 93.7 -> 94.4; 93.8 -> 94.7; 94.5 -> 94.5 | Mixed |
| apply all rules together at each layer (ablation) | Rule-Guided GNN · FB15K-237 v1–v4 | F1 score: 90.1 -> 83.4; 94.4 -> 90.9; 94.7 -> 93.1; 94.5 -> 93.0 | Hurt |
| use Rule-Guided GNN instead of INDIGO | inductive link prediction · FB15K-237 v1–v4, NELL-995 v1–v4, INDIGO-BM | F1 Score: 85.1 -> 90.1; 90.8 -> 94.4; 91.8 -> 94.7; 92.2 -> 94.5; 83.7 -> 99.5; 81.2 -> 83.9; 91.6 -> 93.7; 46.4 -> 75.2; 94.1 -> | Helped |
Surprising findings
- Applying all rule types together as guidance lowered F1 on FB15K-237 compared with the current selective rule set, suggesting rule guidance can confuse the model.
- Existing explainable GNN and end-to-end rule learners extract rules with generally low standard confidence, so their plausibility is not verified.
Limitations
- Relation chains (closed-path rules) cannot be encoded and must be extracted by enumeration.
- Only rules of forms R1-R8 (length up to 2) are supported; more complex rules and neural networks are left to future work.
- Applying all rule types together can hurt performance.
Experimental setup
Datasets: FB15K-237 inductive versions v1-v4 NELL-995 inductive versions v1-v4 INDIGO-BM
Benchmarks: FB15K-237 inductive NELL-995 inductive INDIGO-BM
Metrics: precision recall accuracy F1 AUC standard confidence of extracted rules percentage of high-quality rules
Baselines: DRUM NCRL MGNN INDIGO MMDRUM
Builds on: R-GCN MGNN INDIGO DRUM
Closest papers by meaning
- Understanding Expressivity of GNN in Rule Learning (ICLR 2024 · similarity 0.90)
- Neuro-Symbolic Integration for Reasoning and Learning on Knowledge Graphs (AAAI 2024 · similarity 0.88)
- Systematic Relational Reasoning With Epistemic Graph Neural Networks (ICLR 2025 · similarity 0.87)
- GraphTrail: Translating GNN Predictions into Human-Interpretable Logical Rules (NeurIPS 2024 · similarity 0.87)
- Sound Logical Explanations for Mean Aggregation Graph Neural Networks (NeurIPS 2025 · similarity 0.86)
- KnowFormer: Revisiting Transformers for Knowledge Graph Reasoning (ICML 2024 · similarity 0.86)
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.