Research map / Graph learning / Knowledge and temporal graphs
Spatio-Temporal Pivotal Graph Neural Networks 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 STPGNN, a GNN for traffic flow forecasting that identifies pivotal nodes via an aggregation/distribution scoring mechanism and processes them with a pivotal graph convolution in parallel with standard graph convolution on non-pivotal nodes.
Problem
Traffic flow forecasting on sensor networks, where existing GNN methods fail to model complex spatio-temporal dependencies around highly connected pivotal nodes and joint spatio-temporal modeling is costly.
Approach
Scores each node's aggregation and distribution capability from a learned time-lagged similarity matrix, selects top-K pivotal nodes (K = N/5), builds a pivotal graph adjacency for them, and runs a pivotal graph convolution in parallel with a graph-convolution plus linear-unit branch for non-pivotal nodes; complexity is O(T_d K N) rather than O(T_d^2 N^2).
Key results
- STPGNN achieves the best MAE on all seven datasets except MAPE on PEMS-BAY, e.g., 14.37 MAE on PEMS03 (9.05% better than the runner-up) and 13.90 MAE on PEMS08 (6.98% better).
- On TaxiBJ, MAPE improves by 14.15% over the best baseline (26.07 vs. 29.76 for DSTAGNN).
- Training time on PEMS04 drops from 150.37 s/epoch (DSTAGNN) to 32.36 s/epoch for STPGNN, with inference 3.93 s vs. 10.21 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STPGNN instead of the strongest baseline | Traffic forecasting models · PEMS03 | MAE: 9.05 | Helped |
| Use STPGNN instead of the strongest baseline | Traffic forecasting models · PEMS04 | MAE: 6.49 | Helped |
| Use STPGNN instead of the strongest baseline | Traffic forecasting models · PEMS07 | MAE: 4.39 | Helped |
| Use STPGNN instead of the strongest baseline | Traffic forecasting models · PEMS08 | MAPE: 10.32 | Helped |
| Use STPGNN instead of the strongest baseline | Traffic forecasting models · England | MAE: 3.48 | Helped |
| Use STPGNN instead of the strongest baseline | Traffic forecasting models · TaxiBJ | MAPE: 14.15 | Helped |
| Use STPGNN instead of the strongest baseline | Traffic forecasting models · PEMS-BAY | MAPE | Hurt |
| Replace identified pivotal nodes with randomly selected pivotal nodes (ablation) | STPGNN · PEMS08 and England | prediction accuracy | Hurt |
| Remove pivotal graph convolution and use graph convolution plus linear units for all nodes (ablation) | STPGNN · PEMS08 and England | prediction accuracy | Hurt |
| Remove graph convolution and linear units, using only pivotal graph convolution (ablation) | STPGNN · PEMS08 and England | prediction accuracy | Hurt |
| Increase pivotal-node count K toward the reported best setting (ablation) | STPGNN · PEMS07 | MAE: 21.32 -> 20.52 | Helped |
| Increase pivotal-node count K beyond the reported best setting (ablation) | STPGNN · PEMS07 | MAE: 20.52 -> 20.63 | Hurt |
Surprising findings
- Removing the pivotal graph convolution does not uniformly hurt; the effect of the pivotal module varies across datasets (e.g., on England, pivotal convolution matters more than the non-pivotal branch).
Limitations
- Pivotal node count K is a fixed fraction (N/5) chosen empirically rather than learned.
- Pivotal convolution benefit is inconsistent across datasets, per the ablation.
- MAPE on PEMS-BAY is only sub-optimal relative to baseline.
Experimental setup
Datasets: PEMS03 PEMS04 PEMS07 PEMS08 PEMS-BAY England TaxiBJ
Benchmarks: PEMS03 PEMS04 PEMS07 PEMS08 PEMS-BAY England TaxiBJ
Metrics: MAE MAPE RMSE
Baselines: ARIMA DCRNN GWNet STSGCN MTGNN DMSTGCN DSTAGNN TPGNN SGP
Builds on: STSGCN Graph WaveNet DSTAGNN linear unit (gated convolution)
Closest papers by meaning
- Spatial-Temporal Fusion Graph Neural Networks for Traffic Flow Forecasting (AAAI 2021 · similarity 0.92)
- DSTAGNN: Dynamic Spatial-Temporal Aware Graph Neural Network for Traffic Flow Forecasting (ICML 2022 · similarity 0.91)
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.91)
- Spatio-Temporal Graph Neural Point Process for Traffic Congestion Event Prediction (AAAI 2023 · similarity 0.91)
- TESTAM: A Time-Enhanced Spatio-Temporal Attention Model with Mixture of Experts (ICLR 2024 · similarity 0.89)
- Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting (AAAI 2025 · similarity 0.88)
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.