[Paper Review] STGC-GNNs: A GNN-based traffic prediction framework with a spatial-temporal Granger causality graph

Silu He, Qinyao Luo+3 more|arXiv (Cornell University)|Oct 30, 2022

Traffic Prediction and Management Techniques4 citations

TL;DR

This paper proposes STGC-GNNs, a GNN-based traffic prediction framework that models global and dynamic spatial dependence using spatial-temporal Granger causality (STGC) to capture long-range, macroscopic traffic flow transmission. By detecting stable causal relationships in long-term traffic time series via a spatial-temporal alignment algorithm and Granger causality testing, STGC-GNNs improve 45-minute and 1-hour predictions across three backbone models on the METR_LA dataset, outperforming static spatial graphs.

ABSTRACT

The key to traffic prediction is to accurately depict the temporal dynamics of traffic flow traveling in a road network, so it is important to model the spatial dependence of the road network. The essence of spatial dependence is to accurately describe how traffic information transmission is affected by other nodes in the road network, and the GNN-based traffic prediction model, as a benchmark for traffic prediction, has become the most common method for the ability to model spatial dependence by transmitting traffic information with the message passing mechanism. However, existing methods model a local and static spatial dependence, which cannot transmit the global-dynamic traffic information (GDTi) required for long-term prediction. The challenge is the difficulty of detecting the precise transmission of GDTi due to the uncertainty of individual transport, especially for long-term transmission. In this paper, we propose a new hypothesis\: GDTi behaves macroscopically as a transmitting causal relationship (TCR) underlying traffic flow, which remains stable under dynamic changing traffic flow. We further propose spatial-temporal Granger causality (STGC) to express TCR, which models global and dynamic spatial dependence. To model global transmission, we model the causal order and causal lag of TCRs global transmission by a spatial-temporal alignment algorithm. To capture dynamic spatial dependence, we approximate the stable TCR underlying dynamic traffic flow by a Granger causality test. The experimental results on three backbone models show that using STGC to model the spatial dependence has better results than the original model for 45 min and 1 h long-term prediction.

Motivation & Objective

• To address the limitation of existing GNN-based traffic models in capturing global and dynamic traffic information transmission (GDTi) over long horizons.
• To model spatial dependence not as static local connectivity but as a stable, directional causal relationship underlying long-term traffic flow dynamics.
• To develop a method that detects and leverages macroscopic transmitting causal relationships (TCR) in traffic flow using time-series data.
• To improve long-term prediction performance—especially for 45-min and 1-hour forecasts—by replacing static spatial graphs with dynamically inferred causal graphs.
• To validate that STGC-based spatial dependence better captures long-range dependencies than conventional static graph structures.

Proposed method

• Propose a Transmitting Causal Relationship (TCR) hypothesis: although micro-level vehicle behavior is uncertain, macro-level traffic flow exhibits stable causal transmission patterns.
• Develop a spatial-temporal alignment algorithm to identify the causal order and lag between nodes in traffic time series, enabling detection of directional causal relationships.
• Apply the Granger causality test on aligned time series to detect stable causal dependencies, forming a spatial-temporal Granger causality (STGC) graph.
• Integrate the STGC graph as the spatial topology input into GNN-based traffic prediction models, replacing traditional static road network graphs.
• Use three backbone GNN models (e.g., T-GCN, Graph WaveNet, STGCN) to evaluate the STGC-GNN framework under consistent training and inference protocols.
• Train and test the framework on the METR_LA dataset to assess performance across different prediction horizons and node types.

Experimental results

Research questions

• RQ1 Can a stable, directional causal relationship underlie the long-term transmission of global-dynamic traffic information (GDTi) in road networks?
• RQ2 Can spatial-temporal Granger causality effectively detect and model dynamic spatial dependence that is not captured by static spatial graphs?
• RQ3 Does replacing static spatial graphs with STGC-derived graphs improve long-term traffic prediction accuracy?
• RQ4 How does the STGC-GNN framework perform on distant, boundary, and intersection nodes, which are more dependent on long-range dependencies?
• RQ5 Can the STGC-based approach outperform baseline GNN models in 45-minute and 1-hour traffic predictions?

Key findings

• STGC-GNNs significantly outperform baseline GNN models in 45-minute and 60-minute traffic predictions across all three backbone models on the METR_LA dataset.
• The model improves prediction accuracy for all node types, including intersectional, boundary, and distant nodes, demonstrating effective long-range dependency capture.
• The STGC graph captures directional causal relationships that align well with real-world upstream-downstream traffic flow patterns in road networks.
• The spatial-temporal alignment algorithm successfully identifies causal order and lag, enabling accurate detection of TCRs from time series data.
• The Granger causality test applied to aligned time series reveals stable causal dependencies even in dynamically changing traffic conditions.
• The framework shows that modeling spatial dependence as a dynamic, causal structure yields better long-term forecasting performance than static, local connectivity.