Research map / Graph learning / Knowledge and temporal graphs
MINES: Message Intercommunication for Inductive Relation Reasoning over Neighbor-Enhanced Subgraphs
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 MINES, a GraIL-based inductive relation prediction model that inserts undirected GCN layers between RGCN layers for bidirectional message passing and extracts neighbor-enhanced subgraphs including isolated neighbors.
Problem
Inductive relation reasoning on knowledge graphs with GraIL-based models suffers from uni-directional message passing that misses mutual relations and from enclosing subgraph extraction that discards informative neighbors.
Approach
Sequential message intercommunication (RGCN / undirected GCN / RGCN) over a neighbor-enhanced subgraph that adds k-hop isolated neighbors of target entities beyond the enclosing subgraph, so the model recovers omitted mutual message paths and gains discriminative context.
Key results
- MINES achieves AUC-PR of 99.69 on WN18RR v1 and 99.01 on FB15K-237 v1, outperforming prior GraIL-based baselines on most inductive splits.
- On average MINES improves AUC-PR by 4.01% and Hits@10 by 2.78% over the previous best per dataset.
- Ablation shows the message intercommunication mechanism gives larger AUC-PR gains (about 7.7-8.8%) than neighbor-enhanced subgraphs (about 0.3-2.0%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add message intercommunication (M.I.) to GraIL (ablation) | GraIL · WN18RR | AUC-PR: 94.32 -> 99.50 | Helped |
| Add message intercommunication (M.I.) to GraIL (ablation) | GraIL · FB15K-237 | AUC-PR: 84.69 -> 98.46 | Helped |
| Add message intercommunication (M.I.) to GraIL (ablation) | GraIL · NELL-995 | AUC-PR: 86.05 -> 97.16 | Helped |
| Use neighbor-enhanced subgraph (N.E.) instead of enclosing subgraph (ablation) | GraIL · WN18RR | AUC-PR: 94.32 -> 95.68 | Helped |
| Use neighbor-enhanced subgraph (N.E.) instead of enclosing subgraph (ablation) | GraIL · FB15K-237 | AUC-PR: 84.69 -> 85.38 | Helped |
| Use neighbor-enhanced subgraph (N.E.) instead of enclosing subgraph (ablation) | GraIL · NELL-995 | AUC-PR: 86.05 -> 87.80 | Helped |
| Use neighbor-enhanced subgraph (N.E.) instead of enclosing subgraph (ablation) | GraIL · WN18RR | Hit@10: 82.45 -> 87.03 | Helped |
| Use neighbor-enhanced subgraph (N.E.) instead of enclosing subgraph (ablation) | GraIL · FB15K-237 v3 | Hit@10: 91.41 -> 91.41 | No clear effect |
| Use sequential RGCN-GCN-RGCN intercommunication (RGR) (ablation) | MINES · WN18RR v1 | AUC-PR: 95.96 -> 99.69 | Helped |
| Use sequential RGCN-GCN-RGCN intercommunication (RGR) (ablation) | MINES · FB15K-237 v1 | AUC-PR: 83.10 -> 99.01 | Helped |
| Use sequential RGCN-GCN-RGCN intercommunication (RGR) (ablation) | MINES · NELL-995 v1 | AUC-PR: 70.64 -> 99.55 | Helped |
| Replace RGR with GRR (ablation) | MINES · FB15K-237 v1 | AUC-PR: 85.61 -> 99.01 | Helped |
Surprising findings
- Hits@10 improvement on FB15K-237 is small, suggesting gains are larger on sparser datasets with fewer relations.
- Replacing RGCN with GCN in the intercommunication layer reduces parameters by about 1500 per dataset while improving performance.
Limitations
- Gains on FB15K-237 Hits@10 are not significant, suggesting less benefit on denser graphs.
- Uses only structural information; no relation semantics from language models for the unlabeled mutual edges.
Experimental setup
Datasets: WN18RR v1-v4 FB15K-237 v1-v4 NELL-995 v1-v4
Benchmarks: WN18RR v1-v4 FB15K-237 v1-v4 NELL-995 v1-v4
Metrics: AUC-PR Hits@10
Baselines: Neural-LP DRUM RuleN GraIL TACT CoMPILE RPC-IR
Builds on: GraIL RGCN GCN TACT CoMPILE
Closest papers by meaning
- Towards Graph Foundation Models: Training on Knowledge Graphs Enables Transferability to General Graphs (NeurIPS 2025 · similarity 0.86)
- Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.86)
- Stability and Generalization Capability of Subgraph Reasoning Models for Inductive Knowledge Graph Completion (ICML 2025 · similarity 0.85)
- S²DN: Learning to Denoise Unconvincing Knowledge for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.85)
- Talk like a Graph: Encoding Graphs for Large Language Models (ICLR 2024 · similarity 0.84)
- Context-aware Inductive Knowledge Graph Completion with Latent Type Constraints and Subgraph Reasoning (AAAI 2025 · similarity 0.84)
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.