Analysis and visualization of accidents severity based on LightGBM-TPE

In recent years, road traffic accidents, as a leading cause of accidental deaths, have been attracting more and more attention across several disciplines. Notably, the feature study on accidents severity can help exactly identify causality between different risk factors and road accidents, thereby substantially improving road traffic safety. Meanwhile, the application of data visualization to traffic safety investigations is still lacking. Motivated by this, we incorporate the visualization method into machine learning to analyze the traffic accidents data of the UK in 2017. A hybrid algorithm, namely Light Gradient Boosting Machine-Tree-structured Parzen Estimator (LightGBM-TPE) is proposed. Compared with other typical machine learning algorithms, it performs better in terms of the metrics f1,accuracy, recall and precision. Using LightGBM-TPE to calculate the SHAP value of each feature, we find that “Longitude”, “Latitude”, “Hour” and “Day_of_Week” are four risk factors most closely related with accident severity. Visualization for the data further verifies this conclusion. Overall, our research tries to explore an innovative way to understand and evaluate feature importance of road traffic accidents, which can help suggest effective solutions to improve traffic safety. •We propose a hybrid machine learning model, LightGBM-TPE in understanding and predicting traffic accidents.•We utilize data visualization method to investigate feature importance of accident severity.•Our model and method can be adopted to provide practical advices for urban planners.