Research map / Graph learning / Knowledge and temporal graphs
Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting
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 STDN, a dynamic-graph GCN with spatio-temporal embeddings and a trend-seasonality decomposition feeding a GRU encoder and Transformer-style decoder for traffic flow forecasting, and releases the JiNan inner-city traffic dataset.
Problem
Traffic flow forecasting struggles to model complex spatial and temporal dynamics, including long/short periodicities and global node interactions that static distance-based graphs miss.
Approach
Builds a dynamic relationship graph from learnable time-slot, start-node and end-node embeddings with a core tensor, and disentangles node hidden states into trend and seasonal parts via Hadamard interaction with spatio-temporal embeddings (time-of-day, day-of-week, Laplacian eigenvector spatial embeddings) before encoder-decoder prediction.
Key results
- On PeMS04, STDN achieves MAE 18.40, RMSE 30.41, MAPE 12.21%, beating the next best SSTBAN (18.89 / 31.21 / 12.62%).
- On JiNan, STDN achieves MAE 8.32, RMSE 14.52, MAPE 32.19%, versus SSTBAN at 8.41 / 14.62 / 33.79%.
- On PeMS07, STDN reaches MAE 20.08 and RMSE 33.73, while STWave has the best MAPE (8.58%) and STDN's MAPE is 9.29%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add temporal embedding modeling (ablation) | STDN · PeMS04 | MAE: 19.10 -> 18.40 | Helped |
| add temporal embedding modeling (ablation) | STDN · JiNan | MAE: 8.47 -> 8.32 | Helped |
| add spatial embedding modeling (ablation) | STDN · PeMS04 | MAE: 18.55 -> 18.40 | Helped |
| add spatial embedding modeling (ablation) | STDN · JiNan | MAE: 8.44 -> 8.32 | Helped |
| add spatiotemporal embedding (ablation) | STDN · PeMS04 | MAE: 18.93 -> 18.40 | Helped |
| add spatiotemporal embedding (ablation) | STDN · JiNan | MAE: 8.55 -> 8.32 | Helped |
| add dynamic relationship graph learning (ablation) | STDN · PeMS04 | MAE: 18.61 -> 18.40 | Helped |
| add dynamic relationship graph learning (ablation) | STDN · JiNan | MAE: 8.37 -> 8.32 | Helped |
| use spatiotemporal-aware trend-seasonality decomposition instead of Autoformer decomposition (ablation) | STDN · PeMS04 | MAE: 18.60 -> 18.40 | Helped |
| use spatiotemporal-aware trend-seasonality decomposition instead of Autoformer decomposition (ablation) | STDN · JiNan | MAE: 8.75 -> 8.32 | Helped |
| increase decoder layers; performance stabilizes at 2 layers | STDN · PeMS04 and PeMS07 | performance | Mixed |
| vary feature count; use 128 features on PeMS04 and 96 on PeMS07 | STDN · PeMS04 and PeMS07 | performance | Mixed |
Surprising findings
- On PeMS07, STDN does not win on MAPE (9.29% vs STWave 8.58% and STGNCDE 8.80%), so its advantage is not uniform across datasets and metrics.
Limitations
- Advantage is not uniform: STDN is not best on PeMS07 MAPE.
- Reported gains over SSTBAN and STWave are small (e.g., MAE 18.40 vs 18.89 on PeMS04).
Experimental setup
Datasets: PeMS04 PeMS07 JiNan
Benchmarks: PeMS04 PeMS07 JiNan
Metrics: MAE RMSE MAPE
Baselines: HA ARIMA VAR SVR LSTM DCRNN STGCN GWNet GMAN ASTGCN AGCRN DMSTGCN
Builds on: Autoformer Bottleneck Transformer AGCRN GWNet
Closest papers by meaning
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.89)
- Decomposed Spatio-Temporal Mamba for Long-Term Traffic Prediction (AAAI 2025 · similarity 0.89)
- Learning Continuously Evolving Spatio-Temporal Explanations for Traffic Flow Forecasting (NeurIPS 2026 · similarity 0.89)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.89)
- Spatio-Temporal Pivotal Graph Neural Networks for Traffic Flow Forecasting (AAAI 2024 · similarity 0.88)
- TESTAM: A Time-Enhanced Spatio-Temporal Attention Model with Mixture of Experts (ICLR 2024 · 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.