Attention based spatio-temporal graph convolutional network with focal loss for crash risk evaluation on urban road...

Area of Science:

• Transportation Engineering
• Artificial Intelligence
• Data Science

Background:

• Urban road networks face significant crash risks due to complex structures and multi-source data.
• Evaluating crash risk is challenging due to high-dimensional spatio-temporal correlations and imbalanced datasets.

Purpose of the Study:

• To develop and evaluate an advanced model for assessing crash risk in urban road networks.
• To address challenges posed by complex network structures, multi-source data, and data imbalance in crash risk prediction.

Main Methods:

• An Attention based Spatio-Temporal Graph Convolutional Network (ASTGCN) model was developed, incorporating a focal loss function.
• The model utilizes graph convolution to capture spatio-temporal properties and an attention mechanism to identify critical risks.
• Performance was evaluated using real-world urban traffic data and compared against baseline models like ANN, RF, and DSTGCN.

Main Results:

• The ASTGCN model demonstrated superior performance compared to baseline methods in evaluating crash risk.
• The focal loss function significantly improved model performance by addressing dataset imbalance.
• Traffic flow was identified as the most critical factor influencing model performance.

Conclusions:

• The proposed ASTGCN model offers an efficient solution for evaluating dynamic crash risk in urban road networks.
• Findings support enhanced safety management strategies for urban road transportation through accurate risk assessment.