Research map / Graph learning / Knowledge and temporal graphs
Responsive Dynamic Graph Disentanglement for Metro 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 ReDyNet, a dynamic graph neural network for metro flow forecasting that builds responsive latent graphs from flow, date and weather context and disentangles redundancy via an information-bottleneck objective.
Problem
Forecasting metro passenger inflow/outflow in urban rail systems, where flow dependencies are dynamic, context-dependent, and partly redundant, and predefined station graphs fail to capture them.
Approach
Replaces predefined station adjacency with a dynamic adjacency learned in latent space from temporal-aware spatial, date and weather embeddings, uses dynamic Chebyshev graph convolution with a parameter pool, and adds an information-bottleneck redundancy disentanglement module (VAE-style KL term) to separate essential from redundant signals.
Key results
- ReDyNet reports 19.36 MAE / 34.05 RMSE / 18.93% MAPE at 15 min on Beijing Metro, the best among compared baselines (PDFormer 20.96 / 36.12 / 20.50%).
- On Shanghai Metro at 60 min, ReDyNet reports 22.86 MAE vs 23.88 for PDFormer, the strongest baseline shown.
- Ablations show removing the dynamic adjacency or the parameter pool degrades performance most, and removing temporal-aware spatial embedding hurts most among context blocks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ReDyNet instead of PDFormer | Metro flow forecasting · Beijing Metro, Shanghai Metro, and Hangzhou Metro (15min, 30min, 45min, 60min) | MAE: Beijing: 20.96 -> 19.36 (15min), 22.12 -> 20.07 (60min); Shanghai: 22.07 -> 21.15 (15min), 23.88 -> 22.86 (60min); Hangz | Helped |
| add Redundancy Context Disentanglement (ablation) | ReDyNet · Metro flow forecasting | forecasting error | Helped |
| add Task Adaption gating (ablation) | ReDyNet · Metro flow forecasting | forecasting accuracy | Helped |
| use a dynamic adjacency matrix (ablation) | ReDyNet graph learning · Metro flow forecasting | forecasting error | Helped |
| use a graph-convolution parameter pool for dynamic convolution (ablation) | ReDyNet · Metro flow forecasting | forecasting error | Helped |
| add Date Embedding and Weather Embedding (ablation) | ReDyNet Contextual Spatial Embedding · Beijing Metro 15min and 60min | MAE: 19.53 -> 19.36 (15min); 20.66 -> 20.07 (60min) | Helped |
| add Weather Embedding (ablation) | ReDyNet Contextual Spatial Embedding · Beijing Metro 15min and 60min | MAE: 19.42 -> 19.36 (15min); 20.09 -> 20.07 (60min) | Helped |
| add Date Embedding (ablation) | ReDyNet Contextual Spatial Embedding · Beijing Metro 15min and 60min | MAE: 19.81 -> 19.36 (15min); 20.87 -> 20.07 (60min) | Helped |
| add Temporal-aware Spatial Embedding (ablation) | ReDyNet Contextual Spatial Embedding · Beijing Metro 15min and 60min | MAE: 24.80 -> 19.36 (15min); 31.07 -> 20.07 (60min) | Helped |
Surprising findings
- Removing the redundancy disentanglement module has only a slight impact on short-term forecasting, with larger effects at longer horizons.
Limitations
- Only three metro datasets from Chinese cities are evaluated.
- Relies on date and weather context; authors note weather description is coarse and contributes least.
- Hyperparameters chosen by grid search, no statistical significance reported.
Experimental setup
Datasets: Beijing Metro Shanghai Metro Hangzhou Metro
Benchmarks: Beijing Metro Shanghai Metro Hangzhou Metro
Metrics: MAE RMSE MAPE
Baselines: DCRNN STGCN AGCRN STTN FourierGNN ASTGCN GMAN Graph WaveNet STDGRL PDFormer
Builds on: Chebyshev graph convolution information bottleneck AGCRN variational information bottleneck
Closest papers by meaning
- Disentanglement-Guided Spatial-Temporal Graph Neural Network for Metro Flow Forecasting (Student Abstract) (AAAI 2024 · similarity 0.92)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.87)
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.85)
- Topology-aware Neural Flux Prediction Guided by Physics (ICML 2025 · similarity 0.84)
- TelTrans: Applying Multi-Type Telecom Data to Transportation Evaluation and Prediction via Multifaceted Graph Modeling (AAAI 2024 · similarity 0.83)
- Bridging Traffic State and Trajectory for Dynamic Road Network and Trajectory Representation Learning (AAAI 2025 · similarity 0.83)
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.