GTGCN: A gated temporal-graph coupling network model based on contrastive learning for traffic prediction

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Computer Science

Background:

• Accurate traffic prediction is vital for intelligent transportation systems.
• Graph Neural Networks (GNNs) face challenges like over-smoothing and node heterogeneity.
• Existing GNNs struggle with complex spatiotemporal dependencies in traffic data.

Purpose of the Study:

• To propose a novel Gated Temporal-Graph Coupling Network Model (GTGCN) for enhanced traffic prediction.
• To mitigate over-smoothing and improve adaptability to node heterogeneity in GNNs.
• To achieve robust and accurate traffic flow forecasting.

Main Methods:

• A decoupled architecture combining a node-wise Transformer for temporal dependencies and a shallow GCN for spatial aggregation.
• A gated fusion mechanism to integrate spatial information and mitigate over-smoothing.
• Node-level contrastive learning via feature masking and edge dropping to enhance robustness against node heterogeneity.

Main Results:

• GTGCN consistently outperforms ten baseline models across short, medium, and long-term predictions on PEMS04, PEMS08, and METR-LA datasets.
• Achieved significant reductions in Mean Absolute Error (MAE), e.g., 9.6%-17.9% for 60-min predictions compared to AGCRN.
• Ablation studies confirmed the effectiveness of each proposed component.

Conclusions:

• The proposed GTGCN model offers a robust and efficient solution for traffic prediction.
• The gated fusion mechanism adaptively balances node information for heterogeneous traffic networks.
• The decoupled design effectively suppresses the over-smoothing issue in GNNs for traffic forecasting.