Research map / Video / Video segmentation and tracking
Disentanglement-Guided Spatial-Temporal Graph Neural Network for Metro Flow Forecasting (Student Abstract)
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 DGST, a spatial-temporal GNN for metro flow forecasting that disentangles trend and periodicity via variational pre-training and uses dynamic attention-based spatial correlation, outperforming STGCN, ASTGCN, AGCRN, STGODE and MegaCRN on Hangzhou Metro.
Problem
Metro flow forecasting ignores characteristic patterns (trend, periodicity), suffers noise from hand-crafted granularity segments, and uses static adjacency matrices that fail to capture time-varying spatial dependencies between stations.
Approach
Variational disentanglement pre-training separates trend and periodicity latent representations (with a JS-based mutual-information-style estimator), predicts each characteristic separately (MLPs for trend, Fourier-domain attention for periodicity), and uses self-attention over each time step's input plus trend/periodicity similarity adjacency matrices to model dynamic spatial correlations within
Key results
- DGST achieves inflow MAE 22.59 at 15 min vs. 23.24 for the best baseline AGCRN on Hangzhou Metro.
- DGST achieves outflow MAE 24.62 at 60 min vs. 24.52 for AGCRN, so it is not uniformly best on every horizon.
- Ablation (Figure 1) shows the base model, w/o Pre-train, and w/o DySG variants are each worse than full DGST on inflow MAE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, inflow 15 min | MAE: 23.24 -> 22.59 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, inflow 30 min | MAE: 24.09 -> 23.34 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, inflow 60 min | MAE: 26.05 -> 24.71 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, outflow 15 min | MAE: 21.90 -> 21.78 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, outflow 30 min | MAE: 22.90 -> 22.51 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, outflow 60 min | MAE: 24.52 -> 24.62 | Hurt |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, inflow 15 min | RMSE: 39.26 -> 38.56 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, inflow 30 min | RMSE: 41.20 -> 40.30 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, inflow 60 min | RMSE: 46.33 -> 44.40 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, outflow 15 min | RMSE: 39.32 -> 37.11 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, outflow 30 min | RMSE: 40.54 -> 38.23 | Helped |
| Use DGST instead of AGCRN | Metro flow forecasting · Hangzhou Metro, outflow 60 min | RMSE: 43.08 -> 41.84 | Helped |
Surprising findings
- On outflow 60-min MAE, AGCRN (24.52) slightly beats DGST (24.62), contrary to the claim that DGST outperforms all baselines in both short- and long-term forecasts.
Limitations
- Evaluated on a single real-world dataset (Hangzhou Metro).
- Outflow 60-min MAE is slightly worse than AGCRN, and the text overstates superiority.
- Short student abstract gives limited detail on architecture hyperparameters and training setup.
Experimental setup
Datasets: Hangzhou Metro dataset
Benchmarks: Hangzhou Metro dataset
Metrics: MAE RMSE MAPE
Baselines: STGCN ASTGCN AGCRN STGODE MegaCRN
Builds on: ASTGCN MegaCRN Wang et al. 2022 disentanglement
Closest papers by meaning
- Responsive Dynamic Graph Disentanglement for Metro Flow Forecasting (AAAI 2025 · similarity 0.92)
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.87)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.87)
- Spatio-Temporal Pivotal Graph Neural Networks for Traffic Flow Forecasting (AAAI 2024 · similarity 0.86)
- ST-FiT: Inductive Spatial-Temporal Forecasting with Limited Training Data (AAAI 2025 · similarity 0.86)
- SSL-STMFormer Self-Supervised Learning Spatio-Temporal Entanglement Transformer for Traffic Flow Prediction (AAAI 2025 · similarity 0.85)
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.