Predicting road traffic accident severity from imbalanced data using VAE attention and GCN

Traffic accidents have emerged as a significant factor influencing social security concerns. By achieving precise predictions of traffic accident severity, it is conceivable to mitigate the frequency of hazards and enhance the overall safety of road operations. However, since most accident samples are normal cases, only a minority represent major accidents, but the information contained within the minority samples is of utmost importance for accident prediction outcomes. Hence, it is urgent to solve the impact of unbalanced samples on accident prediction. This paper presents a traffic accident severity prediction method based on the Variational Autoencoders (VAE) with self-attention mechanism and Graph Convolutional Networks (GCN) methods. The generation model is established in minority samples by the VAE, and the latent dependence between the accident features is captured by combining with the self-attention mechanism. Since the integer characteristics of the accident samples, the smooth L1 loss function is utilized as the reconstruction error to improve the model optimization ability and generate high-quality data in line with the real characteristics. Considering the interaction relationship between the environmental features of the accident, the graph convolutional networks (GCN) method that combined with the swish function is applied to establish the topological structure, extract the underlying internal relationship, and improve the nonlinear characteristics of the accident prediction model. The real traffic accident samples are used, the experimental results show that the samples generated by the proposed method can improve the accuracy of accident severity by 20%, and serious accidents can be accurately predicted. The results can provide a basis for decision-making for traffic safety systems or managers.