Research map / Graph learning / Knowledge and temporal graphs
Long Range Propagation on Continuous-Time Dynamic Graphs
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 CTAN, an ODE-based continuous-time dynamic graph network built from anti-symmetric weights and forward Euler steps, to propagate information over long spatial and temporal ranges, and introduces long-range benchmarks (temporal path graph, temporal PascalVOC-SP).
Problem
Continuous-time dynamic graph (C-TDG) models perform poorly on long-range tasks that require correlating information from far away in space (higher-order neighbors) or time (past events), because of over-squashing and dissipative dynamics in message passing and RNN memories.
Approach
Use a non-dissipative ODE whose weight matrix is anti-symmetric (minus a small gamma*I for Euler stability), discretized by forward Euler with per-event sub-problems whose initial state carries the node's previous embedding, so information persists over time and propagates spatially as the number of steps grows.
Key results
- On the temporal path graph task, CTAN reaches 88.93% accuracy at n=20 while all baselines stay near chance (about 42-51%) beyond n=9 or 15.
- On temporal PascalVOC-SP (sc=30), CTAN achieves 9.98 to 10.41 F1 across 1-5 GCLs, beating the best baseline TGAT by about 65% with one GCL and 16% with three.
- On future link prediction over Wikipedia, Reddit, LastFM and MOOC, CTAN ranks first on all four, with AUC of 97.55, 98.61, 83.81 and 92.47.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CTAN instead of DyGFormer | Temporal path graph sequence classification · T-PathGraph n=3 | accuracy: 100.0$_{\pm0.0}$ vs 100.0$_{\pm0.0}$ | No clear effect |
| Use CTAN instead of JODIE | Temporal path graph sequence classification · T-PathGraph n=15 | accuracy: 60.0$_{\pm14.91}$ -> 93.47$_{\pm8.78}$ | Helped |
| Use CTAN instead of DyRep | Temporal path graph sequence classification · T-PathGraph n=20 | accuracy: 50.47$_{\pm2.88}$ -> 88.93$_{\pm12.06}$ | Helped |
| Use CTAN instead of DyGFormer | Temporal Pascal VOC classification, 1 GCL · Temporal Pascal VOC sc=10 | F1-score: 8.45$_{\pm0.13}$ -> 7.89$_{\pm0.33}$ | Hurt |
| Use CTAN instead of TGAT | Temporal Pascal VOC classification, 3 GCLs · Temporal Pascal VOC sc=30 | F1-score: 8.79$_{\pm0.29}$ -> 10.16$_{\pm0.52}$ | Helped |
| Use CTAN instead of TGAT | Temporal Pascal VOC classification, 5 GCLs · Temporal Pascal VOC sc=30 | F1-score: 10.38$_{\pm0.7}$ -> 10.41$_{\pm0.52}$ | No clear effect |
| Use CTAN instead of TGN | Future link prediction · Wikipedia | AUC: 95.60$_{\pm0.18}$ -> 97.55$_{\pm0.09}$ | Helped |
| Use CTAN instead of TGN | Future link prediction · Reddit | AUC: 98.23$_{\pm0.10}$ -> 98.61$_{\pm0.04}$ | Helped |
| Use CTAN instead of TGAT | Future link prediction · LastFM | AUC: 81.53$_{\pm0.34}$ -> 83.81$_{\pm0.92}$ | Helped |
| Use CTAN instead of TGN | Future link prediction · MOOC | AUC: 90.74$_{\pm0.99}$ -> 92.47$_{\pm0.78}$ | Helped |
Surprising findings
- EdgeBank with an unbounded time window surpasses CTAN and all baselines on LastFM, which the authors attribute to its access to the full temporal adjacency under a larger effective neighborhood.
- Adding a small gamma*I term to make forward Euler stable makes the ODE slightly dissipative, so the non-dissipativeness is traded against numerical stability.
Limitations
- Forward Euler stability requires a gamma term that makes the dynamics slightly dissipative, so non-dissipativeness is only approximate.
- Link prediction results are on classic datasets that the authors note do not inherently require long-range dependencies; the inductive setting is not directly evaluated.
- EdgeBank outperforms CTAN on LastFM when given unbounded time windows, showing sensitivity to the neighbor sampler size hyperparameter.
Experimental setup
Datasets: Temporal path graph (synthetic, introduced) Temporal PascalVOC-SP (derived from PascalVOC-SP / Pascal VOC 2011) Wikipedia Reddit LastFM MOOC
Benchmarks: Temporal path graph sequence classification Temporal PascalVOC-SP Wikipedia Reddit LastFM MOOC
Metrics: accuracy F1-score AUC
Baselines: DyRep JODIE TGAT TGN GraphMixer DyGFormer EdgeBank
Builds on: A-DGN Temporal Graph Networks (TGN) Graph Neural Diffusion / GDE Long Range Graph Benchmark PascalVOC-SP TGAT
Closest papers by meaning
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.88)
- On Measuring Long-Range Interactions in Graph Neural Networks (ICML 2025 · similarity 0.88)
- State Space Models on Temporal Graphs: A First-Principles Study (NeurIPS 2024 · similarity 0.88)
- A Deep Probabilistic Framework for Continuous Time Dynamic Graph Generation (AAAI 2025 · similarity 0.87)
- Continuous-Time Graph Representation with Sequential Survival Process (AAAI 2024 · similarity 0.87)
- Over-squashing in Spatiotemporal Graph Neural Networks (NeurIPS 2025 · similarity 0.87)
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.