Predicting multiple types of traffic accident severity with explanations: A multi-task deep learning framework

发表期刊：

Safety Science

论文层级：

EI; SCI; Scopus; WAJCI; INSPEC;

论文作者：

Yang Zekun; Zhang Wenping; Feng Juan/[J]Safety ScienceVolume 146, 2022.

论文摘要：

Predicting traffic accident severity is essential for traffic accident prevention and vulnerable road user safety. Furthermore, the explainability of the prediction is crucial for practitioners to extract relevant risk factors and implement corresponding countermeasures. Most extant research ignores the property loss severity of traffic accidents and fails to predict different levels of death and property loss severity. Moreover, while the explainability of traditional models is easy to achieve, an explainable design of deep neural network (DNN) is extremely deficient in existing research. Few attempts that incorporate neural networks suffer from the lack of multiple hidden layers and the negligence of structural information when explaining predictions. In this study, we propose a multi-task DNN framework for predicting different levels of injury, death, and property loss severity. The multi-task and deep learning design enables a comprehensive and precise analysis of traffic accident severity. Unlike many black-box DNN algorithms, our framework could identify key factors that cause the three types of traffic accident severity via layer-wise relevance propagation, which generates explanations based on the structure and weights of DNN. Based on the experiments conducted using Chinese traffic accident data, our proposed model predicts traffic accident severity risks with good accuracy and outperforms state-of-the-art methods. Furthermore, the case studies show that the key factors provided by our framework are more reasonable and informative than the explanations provided by baseline methods. Our model is the first multi-task learning model and the first DNN-based model for traffic accident severity prediction to the best of our knowledge.