Spatiotemporal Convolution Traffic Accident Risk Prediction based on Multi-graph Fusion

Traffic accident risk prediction methods are usually based on static geographical attributes for spatial modeling, ignoring inter-regional functional semantic information, resulting in insufficient ability to capture semantic spatial correlation and dynamic spatio-temporal features. A spatiotemporal convolution traffic accident risk prediction model based on multi-graph fusion is proposed. Firstly, the model constructs spatial relationship maps from multiple perspectives to capture local geospatial dependencies and global semantic attributes. Secondly, the attention mechanism adaptively learns the dynamic representation of temporal and spatial features. Finally, the spatial dependence is captured by the graph convolutional network in the spatiotemporal convolution module, and the time dependence is captured by the gated unit with dilatation causal convolution. In addition, the weighted L2 loss function is used to mitigate the zero-inflation caused by sparse data. Through the validation on two datasets NYC and Chicago, the results show that the model has a large performance advantage in traffic accident risk prediction.