A Data-Driven Framework for Traffic Crash Risk Prediction – Exploiting Multi-Source Heterogeneous Data

Frequent traffic crashes on urban roads seriously threaten public safety and traffic operations. Accurate risk prediction is vital for improving management efficiency and developing intervention measures. This paper proposes a traffic crash risk prediction model integrating multi-source heterogeneous data. It constructs a dynamic spatio-temporal graph network (DTGN) based on edge-aware graph convolutional networks (EGCN) and introduces a dynamic threshold risk stratification mechanism and local crash density (LCD) indicators to alleviate the issue of “zero inflation” in low-frequency areas. The model combines graph convolutional and spatio-temporal convolutional networks to extract multi-dimensional spatio-temporal features and enhances the ability to identify high-risk areas through a weighted loss function. The city is partitioned into hexagonal grid units, and a dynamic adjacency matrix is constructed to capture spatial associations and evolutionary features. Experimental results indicate that DTGN performs effectively in processing multi-source data and extracting key risk features, achieving an accuracy rate of 87% in high-risk area predictions, thereby providing more practical early warning support and decision-making basis for urban traffic safety management.