Research map / Graph learning / Knowledge and temporal graphs
TEILP: Time Prediction over Knowledge Graphs via Logical 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 TEILP, an inductive temporal logical reasoning framework that converts TKGs into a temporal event knowledge graph and learns differentiable random walks plus conditional density functions for time (interval) prediction.
Problem
Time prediction over temporal knowledge graphs (predicting when a fact's interval holds), which embedding-based ranking methods handle poorly because they ignore temporal order and distance and cannot handle unseen timestamps.
Approach
Convert TKGs into a temporal event KG with event and timestamp nodes, run a differentiable temporal random walk to learn logical rules with temporal relations, and attach learned conditional probability density functions (Gaussian/exponential over time gaps) to rules to predict query intervals.
Key results
- On interval datasets, TEILP (rule-split) reaches aeIOU 0.3285 on WIKIDATA12k and 0.2996 on YAGO11k (event-split), outperforming TimePlex (0.2636 and 0.2003).
- On timestamp datasets, TEILP (rule-split) achieves MAE 70.06 on ICEWS14, 774.01 on ICEWS05-15 and 94.45 on GDELT100, versus GBDT's 85.81, 910.16 and 94.92 (ICEWS14 and ICEWS05-15 improved, GDELT100 improved over GBDT as well).
- In future event time forecasting (Table 2), TEILP achieves MAE 24.86 on ICEWS14 versus 29.57 for EvoKG and aeIOU 0.3971 on WIKIDATA12k versus 0.1688 for GBDT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TEILP rather than the strongest listed baseline | Time prediction · YAGO11k | aeIOU: 0.1336 -> 0.2996 | Helped |
| Use TEILP rather than the strongest listed baseline | Time prediction · WIKIDATA12k | aeIOU: 0.2923 -> 0.3285 | Helped |
| Use TEILP rather than the strongest listed baseline | Time prediction · ICEWS14 | MAE: 85.81 -> 70.06 | Helped |
| Use TEILP rather than the strongest listed baseline | Time prediction · ICEWS05-15 | MAE: 910.16 -> 774.01 | Helped |
| Use TEILP rather than the strongest listed baseline | Time prediction · GDELT100 | MAE: 94.92 -> 94.45 | Helped |
| Use duration prediction (-td) with event-split modelling (ablation) | TEILP · YAGO11k | aeIOU: 0.2996 -> 0.2675 | Hurt |
| Use duration prediction (-td) with event-split modelling (ablation) | TEILP · WIKIDATA12k | aeIOU: 0.3260 -> 0.3086 | Hurt |
| Use duration prediction (-td) with rule-split modelling (ablation) | TEILP · YAGO11k | aeIOU: 0.2977 -> 0.2573 | Hurt |
| Use duration prediction (-td) with rule-split modelling (ablation) | TEILP · WIKIDATA12k | aeIOU: 0.3285 -> 0.3228 | Hurt |
| Use rule-split instead of event-split modelling (ablation) | TEILP · YAGO11k | aeIOU: 0.2996 -> 0.2977 | Mixed |
| Use rule-split instead of event-split modelling (ablation) | TEILP · WIKIDATA12k | aeIOU: 0.3260 -> 0.3285 | Helped |
| Use rule-split instead of event-split modelling (ablation) | TEILP · ICEWS14 | MAE: 70.72 -> 70.06 | Helped |
Surprising findings
- The regression-based approach was found to essentially memorize the answer and performs much worse in future event time forecasting.
- Embedding-based temporal point process methods (GHNN, EvoKG) fail when not all previous events are provided in the forecasting setting.
Limitations
- Time prediction is limited to modelling time gaps with predefined density function families (Gaussian/exponential) and uniform timestamp quantization.
- Relies on rule sampling via random walks and maximum rule length limits (3 or 5), which may miss longer dependencies.
- Future work is needed for entity or event attributes that change with time (numerical variables).
Experimental setup
Datasets: WIKIDATA12k YAGO11k ICEWS14 ICEWS05-15 GDELT100
Benchmarks: WIKIDATA12k YAGO11k ICEWS14 ICEWS05-15 GDELT100
Metrics: aeIOU TAC MAE
Baselines: HyTE DE-SimplE TNT-ComplEx TimePlex NeuSTIP GBDT GHNN EvoKG
Builds on: TILP Neural-LP TLogic NeuSTIP
Closest papers by meaning
- TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs (NeurIPS 2024 · similarity 0.89)
- Large Language Models-guided Dynamic Adaptation for Temporal Knowledge Graph Reasoning (NeurIPS 2024 · similarity 0.87)
- Interpreting Temporal Knowledge Graph Reasoning (Student Abstract) (AAAI 2024 · similarity 0.87)
- HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces (AAAI 2024 · similarity 0.86)
- CognTKE: A Cognitive Temporal Knowledge Extrapolation Framework (AAAI 2025 · similarity 0.86)
- Integrating Large Language Models and Möbius Group Transformations for Temporal Knowledge Graph Embedding on the Riemann Sphere (AAAI 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.