Research map / Graph learning / Knowledge and temporal graphs
CognTKE: A Cognitive Temporal Knowledge Extrapolation Framework
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 CognTKE, a dual-process-inspired TKG extrapolation framework that builds a temporal cognitive relation digraph (TCR-Digraph) from global one-hop and local multi-hop historical facts and reasons over it with a global shallow reasoner and a local deep reasoner.
Problem
Future-fact (extrapolation) reasoning on temporal knowledge graphs: path-based methods only capture recent local temporal paths and miss longer historical relations related to the query, while embedding methods lack explainability.
Approach
Combine a query-centric subgraph (TCR-Digraph) that holds global one-hop facts over full history and local multi-hop facts over a recent window, then encode it sequentially with a System 1 (one-hop) reasoner and a System 2 (multi-hop) reasoner using temporal relation embeddings, a QTR-GRU message function and graph attention.
Key results
- CognTKE reaches MRR 46.06 on ICE14 and 35.24 on ICE18, versus 45.78 and 33.50 for the best prior baseline TempValid.
- On ICE05-15 CognTKE achieves MRR 53.13 versus 50.31 for TempValid; on WIKI it reaches MRR 83.21, slightly below TempValid's 83.19 only on Hits@1 tradeoffs, with mixed gains across metrics.
- Zero-shot transfer (e.g., trained on ICE14, tested on ICE18) yields MRR 33.07, close to in-domain baselines, and training on ICE05-15 and testing on ICE14 gives 46.71 MRR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the global shallow reasoner (System1) (ablation) | CognTKE · ICE14, ICE18, and ICE05-15 | MRR: 36.12 -> 46.06; 29.01 -> 35.24; 40.61 -> 53.13 | Helped |
| add the local deep reasoner (System2) (ablation) | CognTKE · ICE14, ICE18, and ICE05-15 | MRR: 44.88 -> 46.06; 33.45 -> 35.24; 44.49 -> 53.13 | Helped |
| add the TR-Component (ablation) | CognTKE · ICE14, ICE18, and ICE05-15 | MRR: 45.44 -> 46.06; 34.01 -> 35.24; 51.49 -> 53.13 | Helped |
| use QTR_GRU instead of addition for message aggregation (ablation) | CognTKE · ICE14, ICE18, and ICE05-15 | MRR: 44.72 -> 46.06; 34.55 -> 35.24; 49.82 -> 53.13 | Helped |
| use CognTKE instead of TempValid | TKG extrapolation · ICE14 | MRR: 45.78 -> 46.06 | Helped |
| use CognTKE instead of TiRGN | TKG extrapolation · ICE18 | MRR: 33.66 -> 35.24 | Helped |
| use CognTKE instead of TempValid | TKG extrapolation · ICE05-15 | MRR: 50.31 -> 53.13 | Helped |
| use CognTKE instead of TempValid | TKG extrapolation · WIKI | MRR: 83.19 -> 83.21 | Helped |
Surprising findings
- Removing the global shallow reasoner (System 1) hurts more than removing the local deep reasoner (System 2), since most queries are answered by one-hop historical facts.
- On ICE05-15 zero-shot transfer from ICE14 or ICE18 is only close to TempValid and worse than training on ICE05-15 itself.
Limitations
- Prediction time is longer than REGCN, CEN and TiRGN on all datasets due to TCR-Digraph construction.
- Mixed results against TempValid on WIKI, where it wins on Hits@3/10 only for TempValid.
- Interpretability claims rest on a theorem and case studies rather than a human evaluation.
Experimental setup
Datasets: ICE14 ICE18 ICE05-15 WIKI
Benchmarks: ICE14 ICE18 ICE05-15 WIKI
Metrics: MRR Hits@1 Hits@3 Hits@10
Baselines: CyGNet TANGO TITer RE-GCN CEN TiRGN TLogic CENET RETIA TECHS TempValid
Builds on: TLogic RE-GCN TiRGN TempValid graph attention network
Closest papers by meaning
- Interpreting Temporal Knowledge Graph Reasoning (Student Abstract) (AAAI 2024 · similarity 0.88)
- TeDS: Joint Learning of Diachronic and Synchronic Perspectives in Quaternion Space for Temporal Knowledge Graph Completion (ICML 2025 · similarity 0.87)
- Large Language Models-guided Dynamic Adaptation for Temporal Knowledge Graph Reasoning (NeurIPS 2024 · similarity 0.87)
- TEILP: Time Prediction over Knowledge Graphs via Logical Reasoning (AAAI 2024 · similarity 0.86)
- Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning (ICLR 2024 · similarity 0.86)
- Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties (NeurIPS 2025 · similarity 0.85)
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.