Risk Prediction of Traffic Accidents Based on Temporal Knowledge Graphs and Enhanced Multi-Graph Attention Networks

The occurrence of traffic accidents is characterized by significant spatiotemporal dependencies and multi-factor coupling. Accurately modeling the dynamic evolution of accident risk has become a critical issue in smart transportation and urban safety management. Although existing methods have made progress in temporal modeling or graph-based modeling, they generally suffer from limitations in modeling multi-source heterogeneous information, insufficient dynamic temporal modeling, and limited semantic structure representation. To address these challenges, this paper proposes a traffic accident risk prediction framework based on temporal knowledge graphs, integrating improved graph embeddings and attention mechanisms to enhance the spatiotemporal modeling capability of the model. First, a temporal traffic knowledge graph is constructed, integrating information about accidents, roads, weather, and environmental factors. Based on this, the TFME-DistMult model is designed, which introduces a time-modified embedding mechanism to effectively capture the evolution paths and temporal dependencies of accident-related entities. Next, an Enhanced Multi-Graph Attention Network (EMG-GAT) is proposed to achieve relationshipaware multi-semantic information extraction and cross-graph fusion modeling, improving both spatial structure modeling and semantic representation capability. Furthermore, a Transformer encoder is incorporated to capture long- and short-term temporal dependencies, enabling precise segment-level accident risk prediction. Experimental evaluation is conducted using a segmentlevel accident risk dataset built from real-world data. The results demonstrate that the proposed model outperforms mainstream benchmark methods across multiple metrics, achieving superior prediction accuracy and generalization ability.