Research map / Graph learning / Knowledge and temporal graphs
Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective
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 a decoupled learning framework (DLF) with a disentangled spatial-temporal GCN (DSTG) that splits temporal dependencies into seasonal and trend parts and updates them alternately at different frequencies for dynamic traffic graph prediction.
Problem
Spatial-temporal graph prediction on dynamic graphs whose structure and node distributions evolve over time, where retraining is costly and continual/online methods need sufficient new data and degrade with new nodes.
Approach
Decouples time dependencies into seasonal (trained rarely on three months of data with masked autoencoding, then frozen) and trend (fine-tuned about every half month) components, and fine-tunes only on a subnet of sampled nodes chosen by Wasserstein-distance evolution scores plus new nodes.
Key results
- On the PeMS dynamic traffic dataset, DLF achieves 28.16 MAE / 43.22 RMSE at 12 horizons, outperforming the best retrained baseline DSTG+AD (28.87 / 43.73) and continual baselines, with DLF training time 45809s vs. 93015s for DSTG+AD.
- On Knowair PM2.5 forecasting, DLF reports MAE 29.37, RMSE 38.31 and CSI 53.05%, versus 30.45 MAE and 51.32% CSI for DSTG+AD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use decoupled training with alternating seasonal and trend updates and representative-node sampling | DSTG · Traffic, all nodes over nine months | 12-horizon MAE: 28.87 -> 28.16 | Helped |
| Use decoupled training with alternating seasonal and trend updates and representative-node sampling | DSTG · Traffic, all nodes over nine months | Time(s): 93014.63 -> 45809.21 | Helped |
| Use decoupled training with alternating seasonal and trend updates and representative-node sampling | DSTG · Traffic, new nodes with neighbors | MAE: 33.36 -> 30.24 | Helped |
| Use decoupled training with alternating seasonal and trend updates and representative-node sampling | DSTG · Knowair | MAE: 30.45 -> 29.37 | Helped |
| Use decoupled training with alternating seasonal and trend updates and representative-node sampling | DSTG · Knowair | CSI: 51.32% -> 53.05% | Helped |
| Add temporal decomposition into seasonal and trend patterns (ablation) | DLF · Traffic, 15-min prediction | MAE | Helped |
| Add masked autoencoder enhancement during seasonal-pattern training (ablation) | DLF · Traffic, 15-min prediction | MAE | Helped |
| Freeze seasonal-pattern weights during short-term fine-tuning (ablation) | DLF · Traffic, 15-min prediction | MAE | Helped |
| Fine-tune with representative nodes beyond only newly added nodes (ablation) | DLF · Traffic, 15-min prediction | MAE | Helped |
Surprising findings
- Continual-learning baselines (DSTG+NN, DSTG+TS, DSTG+SK) underperform the retraining baseline DSTG+AD in the reported PeMS results.
Limitations
- DSTG is a simple spatial-temporal model; authors acknowledge more powerful backbones could improve accuracy.
- Generalization to unseen nodes without historical data is not addressed.
- Dynamic graph is simulated by masking 5% of nodes monthly rather than using real evolving networks.
Experimental setup
Datasets: PeMS (CalTrans) traffic data Knowair
Benchmarks: PeMS (CalTrans) Knowair
Metrics: MAE MAPE RMSE CSI training time (s)
Baselines: TCN STGCN STNN ST-GAM DSTG+AD DSTG+NN DSTG+SK DSTG+TS DSTG-Static
Builds on: TrafficStream STKEC DLinear Masked autoencoders PatchTST
Closest papers by meaning
- DSTAGNN: Dynamic Spatial-Temporal Aware Graph Neural Network for Traffic Flow Forecasting (ICML 2022 · similarity 0.91)
- Meta Dynamic Graph for Traffic Flow Prediction (AAAI 2026 · similarity 0.91)
- Spatial-Temporal Synchronous Graph Convolutional Networks: A New Framework for Spatial-Temporal Network Data Forecasting (AAAI 2020 · similarity 0.90)
- Designing Specialized Two-Dimensional Graph Spectral Filters for Spatial-Temporal Graph Modeling (AAAI 2025 · similarity 0.89)
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.89)
- Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting (AAAI 2025 · similarity 0.89)
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.