Research on Urban Traffic Accident Risk Prediction Method Based on Graph Neural Network

Urban traffic safety is a crucial social issue related to public life safety. Traditional machine learning methods are widely used in traffic accident prediction, but they struggle to effectively capture the complex spatiotemporal dependencies and topological structures in accident data. This study proposes an urban traffic accident risk prediction method based on Graph Neural Networks (GNNs). First, the K-Nearest Neighbors (KNN) algorithm is used to construct the accident graph structure. Second, multiple models including Graph Convolutional Networks (GCN), GraphSAGE, and Spatiotemporal Graph Neural Networks (STGNN) are designed and implemented for comparative research. Finally, GNNExplainer is used to interpret the model prediction results, and a geographic risk map is generated for visual analysis. Experimental results show that the GraphSAGE model achieves the optimal comprehensive performance, with an accuracy of 82.45% and an AUC value of 0.8557, which is significantly better than traditional machine learning models (such as Random Forest and XGBoost). This study provides a high-precision and highly interpretable risk early warning method for urban traffic safety management.