[Paper Review] Spatial-Temporal Fusion Graph Neural Networks for Traffic Flow Forecasting

Mengzhang Li, Zhanxing Zhu|arXiv (Cornell University)|Dec 15, 2020

Traffic Prediction and Management TechniquesEngineering22 references68 citations

TL;DR

STFGNN learns hidden spatial-temporal dependencies by constructing a data-driven temporal graph, fusing it with the spatial graph in a unified fusion graph, and using gated dilated convolutions to capture long-range patterns for traffic flow forecasting.

ABSTRACT

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.

Motivation & Objective

• Motivate and address limitations of using fixed spatial graphs and separate spatial/temporal modeling in traffic forecasting.

Proposed method

• Generate a data-driven temporal graph using fast-DTW to capture temporal similarities.
• Form a Spatial-Temporal Fusion Graph (A_STFG) by integrating Spatial Graph A_SG, Temporal Graph A_TG, and Temporal Connectivity A_TC.
• Use STFGN modules that perform graph multiplications with A_STFG and gated linear units to learn local/global dependencies.
• Incorporate a Gated CNN module with large temporal dilation to broaden receptive fields for long-range dependencies.
• Train with a Huber loss and Adam optimizer; stack multiple STFGN layers to deepen the model.

Experimental results

Research questions

• RQ1 Can a data-driven temporal graph improve learning of spatial-temporal dependencies beyond a given spatial adjacency graph?
• RQ2 Does fusion of spatial, temporal similarity, and temporal connectivity graphs improve forecasting accuracy over existing methods?
• RQ3 Can gated dilated CNNs alongside STFGN modules effectively capture long-range spatial-temporal dependencies?
• RQ4 How does STFGNN perform across multiple real-world traffic datasets compared to baselines?

Key findings

• STFGNN consistently outperforms baselines on four public datasets (PEMS03, PEMS04, PEMS07, PEMS08).
• A data-driven temporal graph (A_TG) constructed via fast-DTW captures hidden temporal correlations not present in the given spatial graph.
• Fusion of A_SG, A_TG, and A_TC into A_STFG enables simultaneous modeling of local and global spatial-temporal dependencies.
• Gated dilated convolution expands temporal receptive fields, aiding long-range dependency learning.
• Ablation studies show larger A_STFG and appropriate sparsity in A_TG improve MAE, MAPE, and RMSE; gated convolution further improves performance.