Motorcycles are more prone to serious crashes than other motor vehicles. To analyze factors affecting crash severity in single motorcycle crashes, five classification algorithms are tested. Additionally, to handle the data imbalance embedded in the crash data from the National Collision Database of Canada is collected. this study proposes seven data preprocessing approaches. To compare the classification performance of difference algorithms, The G-mean (geometric mean) is used. Results indicate that XGBoost and RandomOverSampler is the best combination method, with a G-mean of 0.593, 339% higher than the original model (0.135). The SHAP summary plot reveals that the following features play an important role in classifying injury severity in motorcycle crashes: road safety used, road alignment, traffic control, roadway configuration, road surface, person age, person sex, collision configuration and weather condition. These results are useful to guide government agencies to develop policies or standards to alleviate the severity in motorcycle crashes.

CICTP