Source-linked AI summary
Spatial-Temporal Fusion Graph Neural Networks for Traffic Flow Forecasting
Mengzhang Li, Zhanxing Zhu
TL;DR
Traffic forecasting is limited by incomplete spatial graphs, difficult long-range dependency modeling, and tradeoffs between local and global correlation capture. STFGNN constructs data-driven temporal and fusion graphs and combines them with gated dilated convolutions. Across four real-world datasets, it consistently outperforms baseline models, while the paper notes risks of overfitting and error accumulation in long-range settings.
Problem
Existing forecasting models may miss hidden spatial-temporal dependencies, struggle to capture local and global correlations together, and face overfitting or error accumulation in long-range prediction.
Method
STFGNN constructs a data-driven temporal graph, fuses spatial and temporal graphs, and integrates the fusion module with gated dilated convolutions.
Results
STFGNN consistently outperforms baseline models on four real-world traffic datasets.
Takeaways & Limitations
The framework captures hidden spatial-temporal dependencies while supporting long-range dependency extraction through graph fusion and gated convolutions.
Takeaways & Limitations
Long-range spatial-temporal dependency learning can overfit dynamically changing, noisy real-world data, while autoregressive models may accumulate errors.
Abstract
from arXiv · showhide
Spatial-temporal data forecasting of traffic flow is a challenging task because of complicated spatial dependencies and dynamical trends of temporal pattern between different roads. Existing frameworks typically utilize given spatial adjacency graph and sophisticated mechanisms for modeling spatial and temporal correlations. However, limited representations of given spatial graph structure with incomplete adjacent connections may restrict effective spatial-temporal dependencies learning of those models. To overcome those limitations, our paper proposes Spatial-Temporal Fusion Graph Neural Networks (STFGNN) for traffic flow forecasting. SFTGNN could effectively learn hidden spatial-temporal dependencies by a novel fusion operation of various spatial and temporal graphs, which is generated by a data-driven method. Meanwhile, by integrating this fusion graph module and a novel gated convolution module into a unified layer, SFTGNN could handle long sequences. Experimental results on several public traffic datasets demonstrate that our method achieves state-of-the-art performance consistently than other baselines.
Introduction
Traffic forecasting must model spatial and temporal dependencies that extend beyond explicit road adjacency. STFGNN addresses these gaps with data-driven graph fusion and gated dilated convolutions, and experiments report consistent gains over baselines.
- Traffic forecasting matters for Intelligent Transportation Systems and represents broader location-based spatial-temporal data problems.
- Existing models often rely on given spatial adjacency and miss temporal similarities between distant but functionally related roads.The paper cites roads near business districts as an example of distant nodes sharing rush-hour patterns.
- Learnable adjacency matrices may poorly represent complicated spatial-temporal dependencies, while temporal attention can overfit noisy, dynamically changing data in long-range forecasting.
- Current methods struggle to capture local and global correlations together because recurrent, transformer, CNN, and graph models face efficiency, receptive-field, or information-loss tradeoffs.
- STFGNN combines a data-driven temporal graph and spatial-temporal fusion graph with gated dilated CNNs to capture hidden dependencies across local and long ranges.
- Experiments on four real-world datasets report that STFGNN consistently outperforms baseline models across different traffic characteristics and network sizes.
Related Works
Prior work models spatial-temporal forecasting with graph convolutions, attention, and time-series similarity measures. Dynamic Time Warping supplies a shape-based similarity measure, while temporal graph generation uses selected nearest temporal neighbors.
- Graph convolution networks extend convolution through spectral or spatial-neighbor formulations for graph-based learning.
- Spatial-temporal forecasting methods combine graph-based spatial modeling with recurrent, gated, or attention-based temporal modeling.
- The temporal graph generation algorithm initializes a matrix, compares node-pair temporal distances, connects selected neighbors symmetrically, and returns a weighted temporal graph.
- Time-series similarity methods include timestep-based, shape-based, and change-based approaches.
- Dynamic Time Warping computes an alignment-based distance between two time series using a distance matrix, cost matrix, and warping path.
Preliminaries
The road network is represented as a graph whose nodes correspond to sensors or roads and whose adjacency matrix encodes spatial proximity. Traffic forecasting maps previous observations to future traffic speeds.
- The road network graph contains N sensor or road nodes and a spatial adjacency matrix ASG representing node proximity or distance.
- Each graph observation records d traffic features, such as speed or volume, for every sensor at a time step.
- Traffic forecasting learns a function from previous T speed observations to predict future traffic speeds for correlated sensors.
Networks
STFGNN builds spatial-temporal fusion graphs from spatial, temporal, and temporal-connectivity relations, then processes them with parallel graph modules and gated convolutions. Fast-DTW reduces temporal-graph construction complexity, while stacked modules and pooling capture broader dependencies.
- Overall Framework: Each fusion-graph layer uses parallel STFGN modules alongside a Gated CNN module, with module outputs concatenated before the next layer.The Gated CNN uses parallel 1D dilated convolutions to enlarge the temporal receptive field.
- Temporal Graph Generation: Fast-DTW restricts the warping-path search and reduces temporal-distance computation from O(n^2) to O(Tn).The restriction makes temporal-graph construction applicable to large-scale spatial-temporal data.
- Spatial-Temporal Fusion Graph Construction: STFGNN combines spatial, temporal, and temporal-connectivity graphs into a Spatial-Temporal Fusion Graph for dependency aggregation.The three matrices represent spatial neighbours, nodes with similar temporal patterns, and the same node across proximate time steps.
- Spatial-Temporal Fusion Graph Neural Module: STFGNN modules aggregate graph information through repeated matrix multiplication with the fusion graph, residual connections, gated units, and max pooling.This replaces a regular spectral filter with a simplified matrix-multiplication operation and uses GLU nonlinear activation.
- Overall Framework: Parallel processing and iterative temporal slicing allow STFGN layers to stack while reducing the input time dimension from T to T − K + 1.The stated maximum depth is ⌊T/(K−1)⌋−1 layers.
Experiments
Experiments evaluate STFGNN against established baselines on four public traffic datasets and assess its components through ablations. STFGNN consistently outperforms baselines, while results highlight the roles of fusion-graph size, temporal-graph sparsity, and gated convolution.
- Experimental Setup: One hour of 12 continuous historical traffic-flow steps is used to predict the next hour’s 12 continuous steps.The data are split into training, validation, and test sets at a 6:2:2 ratio.
- Experiment Results: STFGNN consistently and overwhelmingly outperforms baseline models on every evaluated dataset.Table 2 reports performance for 60-minute-ahead prediction on PEMS03, PEMS04, PEMS07, and PEMS08.
- Experiment Results: GraphWaveNet performs relatively poorly on these nonsmooth datasets because it cannot concurrently stack spatial-temporal layers and enlarge 1D-CNN receptive fields.The comparison identifies a difficulty in capturing the relevant long-range structure for these traffic series.
- Experiment Analysis: STSGCN’s local spatial-temporal modules can be disturbed by frequent missing values, while smooth time series can magnify their limited representation ability.Its modules use multiplication operations and extract only local spatial-temporal dependencies.
- Ablation Study: Larger spatial-temporal fusion graphs extract more complicated heterogeneity with fewer stacked layers.The ablation study evaluates configurations on PEMS04 and PEMS08 using MAE, MAPE, and RMSE.
- Ablation Study: Temporal-graph sparsity is an important hyperparameter, while gated convolution improves the long-range learning ability of STFGN modules.The study reports that proper temporal-graph sparsity can support spatial-information-free traffic forecasting and that gated convolution improves performance.
Conclusion
STFGNN combines a data-driven graph with the given spatial graph to capture hidden spatial dependencies in spatial-temporal traffic forecasting. Its STFGN and gated CNN modules jointly learn localized heterogeneity and global homogeneity, and experiments show consistent strong performance.
- Conclusion: STFGNN captures hidden spatial dependencies by fusing a novel data-driven graph with the given spatial graph.The framework is proposed for spatial-temporal traffic data forecasting.
- Conclusion: Integrating the STFGN module with a gated CNN that enlarges temporal receptive fields enables simultaneous learning of localized heterogeneity and global homogeneity.The model stacks the integrated modules to process temporal sequences.
- Conclusion: Detailed experiments and analysis report STFGNN’s consistent strong performance while revealing advantages and defects of previous models.The conclusion connects the empirical results with the comparison of prior approaches.