Research map / Graph learning / Knowledge and temporal graphs
Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion
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 TARGI, a GNN-based inductive KGC method that builds global topology-specific relation graphs (six topological patterns) to embed unseen relations and entities, evaluated on six NELL995/FB15K-237-derived inductive datasets.
Problem
Inductive knowledge graph completion where both entities and relations can be unseen; existing methods mostly handle new entities while relation embeddings are randomly initialized or rely on enclosing subgraph sampling or degree-based relation graphs.
Approach
Construct a global relation graph for each of six relation-pair topological patterns (H-T, T-T, H-H, T-H, PARA, LOOP) over the whole KG, then aggregate relation embeddings with attention over these graphs and use them to update entity embeddings, avoiding degree weighting and subgraph sampling.
Key results
- On fully inductive FB-100, TARGI reaches MRR 0.253 and Hit@10 0.429 versus InGram's 0.228 and 0.373.
- On fully inductive NL-100, TARGI reaches MRR 0.314 and Hit@10 0.489, outperforming InGram's 0.295 and 0.457.
- Ablation shows replacing the relation layer with GAT drops FB-100 MRR from 0.253 to 0.155.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TARGI instead of InGram | Inductive KGC · FB-100 | MRR: 0.228 -> 0.253 | Helped |
| use TARGI instead of InGram | Inductive KGC · NL-100 | MRR: 0.295 -> 0.314 | Helped |
| use TARGI instead of NBFNet | Inductive KGC · FB-50 | MRR: 0.133 -> 0.159 | Helped |
| use TARGI instead of InGram | Inductive KGC · NL-50 | MRR: 0.273 -> 0.314 | Helped |
| use TARGI instead of InGram | Inductive KGC · NELL995-v1 | MRR: 0.688 -> 0.692 | Helped |
| use TARGI instead of NBFNet | Traditional inductive KGC · NL-0 | MRR: 0.258 -> 0.241 | Hurt |
| replace relation-weighted aggregation with mean aggregation (ablation) | TARGI · FB-100 and NL-100 | MRR: 0.244 vs 0.253; 0.279 vs 0.314 | Hurt |
| replace relation-weighted aggregation with sum aggregation (ablation) | TARGI · FB-100 and NL-100 | MRR: 0.182 vs 0.253; 0.211 vs 0.314 | Hurt |
| add the relation embedding layer (ablation) | TARGI · FB-100 and NL-100 | MRR: 0.155 -> 0.253; 0.135 -> 0.314 | Helped |
| use all topological relation graphs rather than only T-H (ablation) | TARGI · FB-100 and NL-100 | MRR: 0.188 -> 0.253; 0.223 -> 0.314 | Helped |
| use all topological relation graphs rather than only H-H (ablation) | TARGI · FB-100 and NL-100 | MRR: 0.216 -> 0.253; 0.263 -> 0.314 | Helped |
| use all topological relation graphs rather than only H-T (ablation) | TARGI · FB-100 and NL-100 | MRR: 0.177 -> 0.253; 0.259 -> 0.314 | Helped |
Surprising findings
- On traditional inductive NL-0 and NELL995-v1, TARGI is competitive but not uniformly best, with NBFNet top on NELL995-v1 MRR (0.611 vs TARGI 0.692 is actually higher, so TARGI leads there)
Limitations
- GraIL, CoMPILE, and RMPI could not be run on FB datasets due to scalability issues, so comparisons are incomplete there
- Baseline numbers were reproduced with a uniform embedding dimension of 32, so results may differ from originally reported values
- Pre-training-based methods like GreenKGC and ULTRA are not compared under their original settings
Experimental setup
Datasets: NELL995 FB15K-237 NL-0 NELL995-v1 NL-50 FB-50 NL-100 FB-100
Benchmarks: NELL995-v1 NL-0 NL-50 FB-50 NL-100 FB-100
Metrics: MRR Hit@10 Hit@1 MR
Baselines: GraIL CoMPILE RMPI CompGCN NodePiece NeuralLP DRUM BLP QBLP NBFNet RED-GNN RAILD
Builds on: GATv2 InGram TACT GraIL RMPI
Closest papers by meaning
- InGram: Inductive Knowledge Graph Embedding via Relation Graphs (ICML 2023 · similarity 0.93)
- S²DN: Learning to Denoise Unconvincing Knowledge for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.92)
- Replacing Paths with Connection-Biased Attention for Knowledge Graph Completion (AAAI 2025 · similarity 0.91)
- Context-aware Inductive Knowledge Graph Completion with Latent Type Constraints and Subgraph Reasoning (AAAI 2025 · similarity 0.90)
- KGDM: A Diffusion Model to Capture Multiple Relation Semantics for Knowledge Graph Embedding (AAAI 2024 · similarity 0.89)
- Stability and Generalization Capability of Subgraph Reasoning Models for Inductive Knowledge Graph Completion (ICML 2025 · similarity 0.88)
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.