DualRisk: A Two-Branch Model for Sparse Traffic Accident Risk Forecasting

Intelligent Transportation Systems (ITS) face critical challenges in mitigating the risks posed by traffic accidents, necessitating accurate traffic accident risk forecasting (TARF) to enhance safety and operational efficiency. Existing methods employ Spatiotemporal Graph Neural Networks (STGNNs) to predict traffic accident risks based on traffic dynamics. To address the sparsity of traffic accidents in TARF, these methods employ specialized data transformations or cost-sensitive loss functions, which require manual adjustments to parameters. Additionally, they fail to consider the asynchronicity between the occurrence probability and the potential severity when estimating risks. In this study, we propose DualRisk, a novel TARF model motivated by the observed asynchronous relationship between traffic accident occurrences and severity. DualRisk comprises a spatiotemporal encoding module that captures both recurring and non-recurring traffic patterns, and a two-branch prediction module that simultaneously estimates the occurrence probability and potential severity of accidents. We specifically address the sparsity issue by estimating zero-inflated distributions in the occurrence probability branch, thereby eliminating the need for manual parameter adjustments. Our approach demonstrates consistently superior performance, with the final results averaged over multiple experimental runs. Compared to the suboptimal method, our model reduces the average mean absolute error by 4.26% and improves the Recall metric by 2.86%. Additionally, case studies with visualizations illustrate DualRisk’s capability to capture the asynchronicity in TARF, underscoring its practical application value in real-world environments.