Research map / Graph learning / Knowledge and temporal graphs
TeDS: Joint Learning of Diachronic and Synchronic Perspectives in Quaternion Space for Temporal 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 TeDS, a quaternion-based temporal knowledge graph completion model that separately models diachronic (trigonometric timestamp rotation of relations) and synchronic (composite quaternion interaction) temporal perspectives, reporting SOTA on six TKGC benchmarks.
Problem
Temporal knowledge graph completion typically treats time as supplementary information and does not capture how facts vary across time periods (diachronicity) or how relations interact within a given temporal context (synchronicity).
Approach
Splits timestamps into diachronic and synchronic components within quaternion space: synchronic perception builds composite quaternions from timestamp and relation and applies Hamilton products; diachronic perception maps each time point to an angle-based unit quaternion via cosine/sine and rotates relations with it; the two perceptions are summed in a scoring function.
Key results
- TeDS reports MRR 90.7 on ICEWS14 and 77.3 on YAGO11k, versus TPComplEx 89.8 on ICEWS14 and SANe 25.0 on YAGO11k.
- TeDS reports MRR 95.4 on Wikidata12k and 71.5 on GDELT, outperforming SANe (43.2) and MvTuckER (54.9) respectively.
- Ablation shows SP alone reaches 79.6 MRR on ICEWS14 and DP alone 71.1, both below the full TeDS 90.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TeDS instead of TPComplEx | TKGC models · ICEWS14 | MRR: 89.8 -> 90.7 | Helped |
| use TeDS instead of TPComplEx | TKGC models · ICEWS05-15 | MRR: 84.5 -> 87.3 | Helped |
| use TeDS instead of DLGR | TKGC models · ICEWS18 | MRR: 35.4 -> 45.8 | Helped |
| use TeDS instead of MDRQS | TKGC models · YAGO11k | MRR: 27.1 -> 77.3 | Helped |
| use TeDS instead of SANe | TKGC models · Wikidata12k | MRR: 43.2 -> 95.4 | Helped |
| use TeDS instead of GLARGCN | TKGC models · GDELT | MRR: 57.1 -> 71.5 | Helped |
| add synchronic perception (SP) to DP (ablation) | TeDS · ICEWS14 | MRR: 71.1 -> 90.7 | Helped |
| add diachronic perception (DP) to SP (ablation) | TeDS · ICEWS14 | MRR: 79.6 -> 90.7 | Helped |
| use SP instead of a regular quaternion model (HTM) (ablation) | TeDS · ICEWS14 | MRR: 56.4 -> 90.7 | Helped |
| add synchronic perception (SP) to DP (ablation) | TeDS · Wikidata12k | MRR: 62.8 -> 95.4 | Helped |
| add diachronic perception (DP) to SP (ablation) | TeDS · Wikidata12k | MRR: 67.9 -> 95.4 | Helped |
| use SP instead of a regular quaternion model (HTM) (ablation) | TeDS · Wikidata12k | MRR: 30.2 -> 95.4 | Helped |
Surprising findings
- The paper reports that SP alone outperforms DP alone across datasets, which the authors attribute to most facts concentrating in specific temporal contexts.
Limitations
- Relies on quaternion embeddings with no explicit handling of periodic temporal patterns, which the authors flag as future work.
- Results on some benchmarks (e.g., ICEWS14 MRR near 90.7) are much higher than prior reported numbers, with reported gains that are very large relative to baselines and not independently reproduced here.
- Some analysis (e.g., temporal regularizer and time-constrained bias) is described loosely and the timestamp handling for interval facts is an averaging heuristic.
Experimental setup
Datasets: ICEWS14 ICEWS05-15 ICEWS18 YAGO11k Wikidata12k GDELT
Benchmarks: ICEWS14 ICEWS05-15 ICEWS18 YAGO11k Wikidata12k GDELT
Metrics: MRR Hits@1 Hits@3 Hits@10
Baselines: TTransE TNTComplEx TeMP-SA TeLM RotateQVS TBDRI BTDG TPComplEx CEC-BD SANe MTE MDRQS
Builds on: RotatE QuatE TComplEx TeLM Tucker-style 4th-order tensor factorization
Closest papers by meaning
- HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces (AAAI 2024 · similarity 0.88)
- CognTKE: A Cognitive Temporal Knowledge Extrapolation Framework (AAAI 2025 · similarity 0.87)
- ChroKnowledge: Unveiling Chronological Knowledge of Language Models in Multiple Domains (ICLR 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)
- DanmakuTPPBench: A Multi-modal Benchmark for Temporal Point Process Modeling and Understanding (NeurIPS 2025 · similarity 0.85)
- TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs (NeurIPS 2024 · 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.