Abstract:

To explore the factors influencing the injury severity of cyclists in urban bicycle accidents and mitigate the impact of data heterogeneity and imbalance on the quantification of these factors, this study proposes a method integrating resampling, latent class analysis (LCA), and Bayesian networks (BNs) based on 3 895 bicycle accidents from the CRSS database. First, LCA was used to reclassify accident data into several sub-accident clusters with intra-cluster homogeneity and inter-cluster heterogeneity to reduce the impact of data heterogeneity. Second, random over-sampling (ROS), synthetic minority oversampling technique, and adaptive synthetic sampling approach were used to resample each accident cluster to reduce the impact of data imbalance. Finally, based on various resampled accident clusters, two BN structure learning algorithms and one parameter learning algorithm were applied and the optimal BN model for each accident cluster was selected based on AUC values to enable quantitative and heterogeneity analyses of factors influencing the injury severity of cyclists. Results show that when the overall accident data were divided into three homogeneous sub-clusters, the LCA model achieved an increased entropy value of 0.943. For the C1, C2, C3, and OD accident clusters, 10, 13, 9, and 12 key factors influencing the injury severity of cyclists were identified, respectively. The introduction of LCA and resampling into the BN considerably improved the BN model’s G-mean value, AUC value, and risk factor identification capabilities. Factors such as time period, cyclist’s gender, cyclist’s age, and weather conditions showed substantial heterogeneity across different accident clusters.

Key words: traffic safety, bicycle accidents, injury severity, latent class analysis, Bayesian networks