Research map / Large language models / Retrieval and knowledge
Context-aware Inductive Knowledge Graph Completion with Latent Type Constraints and Subgraph 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 CATS, which fine-tunes Qwen2-7B-Instruct with LoRA using two prompt-based modules (latent type-aware reasoning and subgraph reasoning over degree-filtered paths and neighboring facts) to judge triple plausibility for inductive KG completion.
Problem
Inductive knowledge graph completion must predict missing triples involving unseen entities; existing path-based and GNN/embedding methods depend on reasoning paths or retraining, and prior LLM methods only re-rank outputs of other models or leak validation triples.
Approach
Use an LLM, supervised fine-tuned with contrastive negatives, to judge triple existence from two contexts: latent entity-type consistency inferred from same-relation triples, and a subgraph containing degree-filtered reasoning paths plus embedding-retrieved neighboring facts; the two scores are averaged for ranking.
Key results
- CATS (full) with Qwen2-7B reaches inductive FB15k-237 MRR 0.882 vs. 0.764 for previous SOTA APST, and inductive WN18RR MRR 0.982 vs. 0.908.
- CATS outperforms prior SOTA in 16 of 18 transductive, inductive, and few-shot settings, with average absolute MRR improvement of 7.2%.
- Using Qwen2-1.5B instead of 7B cuts average inference time from 1.43s to 0.51s per sample while still beating previous SOTA in most settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the TAR and SR modules to triple-only fine-tuning (ablation) | CATS · WN18RR, FB15k-237, and NELL-995; transductive and inductive | MRR: 0.947 -> 0.978; 0.800 -> 0.843; 0.811 -> 0.885; 0.901 -> 0.982; 0.814 -> 0.882; 0.710 -> 0.861 | Helped |
| Add supervised fine-tuning (ablation) | CATS with Qwen2-7B · WN18RR, FB15k-237, and NELL-995; transductive and inductive | MRR: 0.225 -> 0.978; 0.197 -> 0.843; 0.177 -> 0.885; 0.207 -> 0.982; 0.179 -> 0.882; 0.208 -> 0.861 | Helped |
| Add TAR and SR guidance rather than relying on the LLM's internal knowledge alone (ablation) | CATS with Qwen2-7B · WN18RR, FB15k-237, and NELL-995; transductive and inductive | MRR: 0.199 -> 0.978; 0.140 -> 0.843; 0.156 -> 0.885; 0.174 -> 0.982; 0.144 -> 0.882; 0.138 -> 0.861 | Helped |
| Combine prompts from TAR and SR rather than using the separated full configuration (ablation) | CATS with Qwen2-7B · WN18RR, FB15k-237, and NELL-995; transductive and inductive | MRR: 0.967 vs 0.978; 0.830 vs 0.843; 0.861 vs 0.885; 0.985 vs 0.982; 0.859 vs 0.882; 0.885 vs 0.861 | Mixed |
| Add neighboring facts to reasoning-path evidence (ablation) | CATS (SR) · WN18RR, FB15k-237, and NELL-995; transductive and inductive | MRR: 0.960 -> 0.972; 0.800 -> 0.829; 0.773 -> 0.869; 0.968 -> 0.992; 0.851 -> 0.875; 0.829 -> 0.906 | Helped |
| Add reasoning paths to neighboring-fact evidence (ablation) | CATS (SR) · WN18RR, FB15k-237, and NELL-995; transductive and inductive | MRR: 0.962 -> 0.972; 0.824 -> 0.829; 0.861 -> 0.869; 0.971 -> 0.992; 0.873 -> 0.875; 0.895 -> 0.906 | Helped |
| Apply degree-based filtering to reasoning paths (ablation) | CATS (SR) · WN18RR, FB15k-237, and NELL-995; transductive and inductive | MRR: 0.960 -> 0.972; 0.801 -> 0.829; 0.769 -> 0.869; 0.968 -> 0.992; 0.844 -> 0.875; 0.821 -> 0.906 | Helped |
| Use separated TAR and SR modules rather than a combined prompt (ablation) | CATS with Qwen2-7B · WN18RR, FB15k-237, and NELL-995; transductive and inductive | Hits@1: 0.941 -> 0.962; 0.750 -> 0.776; 0.779 -> 0.820; 0.971 -> 0.965; 0.778 -> 0.805; 0.811 -> 0.783 | Mixed |
| Use Qwen2-1.5B instead of Qwen2-7B (ablation) | CATS (full) · WN18RR, FB15k-237, and NELL-995; transductive and inductive | MRR: 0.978 -> 0.966; 0.843 -> 0.804; 0.885 -> 0.877; 0.982 -> 0.948; 0.882 -> 0.814; 0.861 -> 0.835 | Mixed |
| Use Llama-3-8B instead of Qwen2-7B (ablation) | CATS (full) · WN18RR, FB15k-237, and NELL-995; transductive and inductive | MRR: 0.978 -> 0.965; 0.843 -> 0.802; 0.885 -> 0.867; 0.982 -> 0.956; 0.882 -> 0.862; 0.861 -> 0.837 | Mixed |
Surprising findings
- Without SFT the LLM performs very poorly (e.g., WN18RR transductive MRR 0.225), showing a large semantic gap between KG triples and natural language prompts.
- Zero-shot LLM judgments with no KG context perform near chance-level (MRR about 0.14-0.20), indicating gains are not from internal LLM knowledge.
Limitations
- Performance depends on the availability and quality of neighboring facts and paths; few-shot transductive NELL-995 is weaker than some baselines.
- Only Qwen2-7B, Llama-3-8B and Qwen2-1.5B were tried; larger LLMs were not explored due to resource constraints.
- Evaluation uses 1 ground truth plus 49 negatives per query, a fixed candidate protocol.
Experimental setup
Datasets: WN18RR FB15k-237 NELL-995
Benchmarks: WN18RR FB15k-237 NELL-995
Metrics: MRR Hits@1
Baselines: RuleN GRAIL RED-GNN Adaprop MINERVA TuckER KG-BERT BERTRL KRST APST
Builds on: APST KRST BERTRL GraIL RED-GNN
Closest papers by meaning
- Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.90)
- S²DN: Learning to Denoise Unconvincing Knowledge for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.90)
- Replacing Paths with Connection-Biased Attention for Knowledge Graph Completion (AAAI 2025 · similarity 0.89)
- Stability and Generalization Capability of Subgraph Reasoning Models for Inductive Knowledge Graph Completion (ICML 2025 · similarity 0.89)
- Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models (ICML 2025 · similarity 0.89)
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.89)
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.