Research map / Graph learning / Knowledge and temporal graphs
Interpreting Temporal Knowledge Graph Reasoning (Student Abstract)
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 LSGAT, an attention-based relation-aware GNN for temporal knowledge graph reasoning that combines short- and long-term dependencies and uses a GNNExplainer-style mask to surface influential historical events as explanations.
Problem
Temporal knowledge graph extrapolation (entity/link prediction) is accurate but GNN-based TKG reasoners are black boxes, lacking interpretability of which historical events drive predictions.
Approach
Combines a hierarchical relational graph attention network (HRGAT) for short-term evolution and R-GCN over a global graph with GRU for long-term dependencies, plus an interpreting module that learns an adjacency mask over the computation graph to identify the most influential historical events for each prediction.
Key results
- LSGAT achieves 30.06 MRR, 19.86 Hits@1 and 50.24 Hits@10 on ICEWS18, versus 29.27 MRR for the best baseline HGLS.
- On ICEWS05-15, LSGAT reaches 46.92 MRR and 68.04 Hits@10, outperforming HGLS (46.19 MRR, 67.15 Hits@10).
- A case study on ICEWS18 lists the most influential historical events for the predicted event (Government(Russia), Accuse, Israel, 9.19), with weights such as 0.35 for (Putin, Make statement, Israel, 9.18).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add short-term dependency to the long-term-only variant (ablation) | LSGAT · ICEWS18 | MRR: 27.01 -> 30.06 | Helped |
| Add short-term dependency to the long-term-only variant (ablation) | LSGAT · ICEWS18 | Hits@1: 17.45 -> 19.86 | Helped |
| Add short-term dependency to the long-term-only variant (ablation) | LSGAT · ICEWS18 | Hits@10: 45.83 -> 50.24 | Helped |
| Add short-term dependency to the long-term-only variant (ablation) | LSGAT · ICEWS05-15 | MRR: 39.44 -> 46.92 | Helped |
| Add short-term dependency to the long-term-only variant (ablation) | LSGAT · ICEWS05-15 | Hits@1: 28.45 -> 35.71 | Helped |
| Add short-term dependency to the long-term-only variant (ablation) | LSGAT · ICEWS05-15 | Hits@10: 61.28 -> 68.04 | Helped |
| Add long-term dependency to the short-term-only variant (ablation) | LSGAT · ICEWS18 | MRR: 29.24 -> 30.06 | Helped |
| Add long-term dependency to the short-term-only variant (ablation) | LSGAT · ICEWS18 | Hits@1: 19.10 -> 19.86 | Helped |
| Add long-term dependency to the short-term-only variant (ablation) | LSGAT · ICEWS18 | Hits@10: 49.35 -> 50.24 | Helped |
| Add long-term dependency to the short-term-only variant (ablation) | LSGAT · ICEWS05-15 | MRR: 45.34 -> 46.92 | Helped |
| Add long-term dependency to the short-term-only variant (ablation) | LSGAT · ICEWS05-15 | Hits@1: 34.05 -> 35.71 | Helped |
| Add long-term dependency to the short-term-only variant (ablation) | LSGAT · ICEWS05-15 | Hits@10: 66.62 -> 68.04 | Helped |
Limitations
- Interpretability is only illustrated with a single qualitative case study; no quantitative explanation evaluation is reported.
- Gains over HGLS on ICEWS18 are small (about 0.8 MRR).
- Hardware, training details and hyperparameters are not stated in the abstract-length paper.
Experimental setup
Datasets: ICEWS18 ICEWS05-15
Benchmarks: ICEWS18 ICEWS05-15
Metrics: MRR Hits@1 Hits@10
Baselines: RE-GCN CyGNet CENET HGLS LSGAT-L LSGAT-S
Builds on: RE-GCN GNNExplainer ConvTransE R-GCN
Closest papers by meaning
- CognTKE: A Cognitive Temporal Knowledge Extrapolation Framework (AAAI 2025 · similarity 0.88)
- TEILP: Time Prediction over Knowledge Graphs via Logical Reasoning (AAAI 2024 · similarity 0.87)
- Large Language Models-guided Dynamic Adaptation for Temporal Knowledge Graph Reasoning (NeurIPS 2024 · similarity 0.86)
- Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning (ICLR 2024 · similarity 0.86)
- TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs (NeurIPS 2024 · similarity 0.85)
- 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.