Research map / Graph learning / Knowledge and temporal graphs
DPCL-Diff：Temporal Knowledge Graph Reasoning Based on Graph Node Diffusion Model with Dual-Domain Periodic Contrastive Learning
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 DPCL-Diff, a temporal knowledge graph reasoning model combining a graph node diffusion model (GNDiff) for new events with dual-domain periodic contrastive learning that maps periodic entities to Poincaré space and non-periodic entities to Euclidean space.
Problem
Predicting future missing facts in temporal knowledge graphs, especially for new events with sparse historical interactions and for periodic events whose candidate entities have very similar representations.
Approach
Applies a discrete masked graph-node diffusion process to generate samples for sparsely related (new) events, and performs contrastive learning across two geometric domains, Poincaré for periodic entities and Euclidean for non-periodic ones, to separate similar periodic candidates; final predictions average the diffusion and DPCL probabilities.
Key results
- On ICEWS14, DPCL-Diff reports MRR 66.59 versus CENET's 51.93, a large gain on the time-filtered benchmark.
- On YAGO, DPCL-Diff reaches MRR 84.45 and Hits@1 84.23, versus RLGNet's MRR 80.17.
- On ICEWS18, DPCL-Diff reaches MRR 47.02, slightly above CyGNet's 46.69; ablation shows removing GNDiff drops ICEWS18 MRR to 44.93.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DPCL-Diff instead of the strongest reported baseline | Temporal knowledge graph reasoning · ICEWS18 | MRR: 46.69 -> 47.02 | Helped |
| Use DPCL-Diff instead of the strongest reported baseline | Temporal knowledge graph reasoning · ICEWS14 | MRR: 51.93 -> 66.59 | Helped |
| Use DPCL-Diff instead of the strongest reported baseline | Temporal knowledge graph reasoning · WIKI | MRR: 66.87 -> 68.44 | Helped |
| Use DPCL-Diff instead of the strongest reported baseline | Temporal knowledge graph reasoning · YAGO | MRR: 80.17 -> 84.45 | Helped |
| Add GNDiff (ablation) | DPCL-Diff · ICEWS18 | MRR: 44.93 -> 47.02 | Helped |
| Add GNDiff (ablation) | DPCL-Diff · YAGO | MRR: 81.30 -> 84.45 | Helped |
| Add DPCL (ablation) | DPCL-Diff · ICEWS18 | MRR: 46.51 -> 47.02 | Helped |
| Add DPCL (ablation) | DPCL-Diff · YAGO | MRR: 84.21 -> 84.45 | Helped |
| Map periodic entities to Poincaré space and non-periodic entities to Euclidean space rather than using the same space for both (ablation) | DPCL · YAGO (without GNDiff) | MRR: 79.64 -> 81.30 | Helped |
| Use GNDiff rather than DPCL for new-event reasoning | New-event reasoning · ICEWS14, ICEWS18, WIKI, and YAGO (new events) | MRR | Helped |
Surprising findings
- Hits@10 on YAGO is slightly lower than RLGNet (84.75 vs 84.96), and on WIKI the gain over CENET is small (MRR 68.44 vs 66.87).
- The reported 30.37% Hits@1 improvement over CENET on ICEWS14 is stated as a main result, while the table lists the Improve row as relative to CENET.
Limitations
- Evaluation on new events is shown on separately constructed subsets, and gains on periodic-heavy datasets like YAGO and WIKI are less pronounced.
- Batch size 64 is cited by the authors as potentially constraining performance.
- Several hyperparameters (alpha, lambda) show different optimal values across datasets.
Experimental setup
Datasets: ICEWS14 ICEWS18 WIKI YAGO
Benchmarks: ICEWS14 ICEWS18 WIKI YAGO
Metrics: MRR Hits@1 Hits@3 Hits@10
Baselines: ComplEx R-GCN ConvE RE-Net xERTE CyGNet EvoKG RPC CENET DHE-TKG RLGNet HisRES
Builds on: CENET discrete diffusion for text supervised contrastive learning
Closest papers by meaning
- Large Language Models-guided Dynamic Adaptation for Temporal Knowledge Graph Reasoning (NeurIPS 2024 · similarity 0.87)
- DanmakuTPPBench: A Multi-modal Benchmark for Temporal Point Process Modeling and Understanding (NeurIPS 2025 · similarity 0.85)
- TEILP: Time Prediction over Knowledge Graphs via Logical Reasoning (AAAI 2024 · similarity 0.84)
- A Selective Learning Method for Temporal Graph Continual Learning (ICML 2025 · similarity 0.84)
- LLM-DR: A Novel LLM-Aided Diffusion Model for Rule Generation on Temporal Knowledge Graphs (AAAI 2025 · similarity 0.83)
- Integrating Large Language Models and Möbius Group Transformations for Temporal Knowledge Graph Embedding on the Riemann Sphere (AAAI 2025 · similarity 0.83)
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.