Predicting Spatiotemporal Traffic Accident Risks with Regression and Classification Integration

Traffic accident risk forecasting is vital for enhancing the safety of Intelligent Transportation Systems (ITS). While current approaches rely on Spatiotemporal Graph Neural Networks (STGNNs) with regression loss functions to encode historical traffic dynamics and predict future accident risks, their effectiveness is limited by the rare occurrence of accidents. To overcome this challenge, we present ClassiRisk, a novel approach that addresses traffic accident risk modeling from both regression and classification perspectives. Our method initiates with the encoding of historical traffic dynamics, incorporating a temporal attention mechanism and a novel fused spatial encoding module to capture both recurring and non-recurring spatiotemporal patterns. Subsequently, the model utilizes these encoded embeddings to predict: 1) a sparse binary mask indicating accident occurrence, employing a weighted classification loss to tackle the zero-inflation problem; and 2) the associated risk levels of accidents. The final prediction of traffic accident risks is derived by amalgamating these two outputs. Experimental evaluations conducted on real-world data validate the superiority of our approach over existing methodologies.