GATR: A Road Network Traffic Violation Prediction Method Based on Graph Attention Network.

Area of Science:

• Transportation Engineering
• Artificial Intelligence
• Data Science

Background:

• Traffic violation prediction is crucial for transportation safety.
• Deep learning methods are emerging for traffic violation prediction.
• Existing spatial grid methods have limitations in expressing spatial correlations with road networks.

Purpose of the Study:

• To propose a novel Graph Attention Network based on Road network (GATR) model for accurate spatiotemporal prediction of traffic violations.
• To improve traffic violation prediction accuracy by leveraging road network topology and diverse features.

Main Methods:

• Developed a GATR model using a graph attention network architecture.
• Integrated historical traffic violation data, environmental features, and urban functional features.
• Employed GNN Explainer for model interpretability and verification.

Main Results:

• The GATR model demonstrated higher prediction accuracy (RMSE = 1.7078) compared to Conv-LSTM (RMSE = 1.9180).
• GATR effectively captures the spatiotemporal distribution patterns of traffic violations.
• Model verification confirmed the rationality of GATR by analyzing road network subgraphs and feature influences.

Conclusions:

• The GATR model offers a more accurate and interpretable approach to traffic violation prediction.
• This method can provide valuable insights for traffic violation prevention and control strategies.
• Implementing GATR can significantly contribute to improving overall traffic safety.