Prediction in Traffic Accident Duration Based on Heterogeneous Ensemble Learning

Abstract

Based on millions of traffic accident data in the United States, we build an accident duration prediction model based on heterogeneous ensemble learning to study the problem of accident duration prediction in the initial stage of the accident. First, we focus on the earlier stage of the accident development, and select some effective information from five aspects of traffic, location, weather, points of interest and time attribute. Then, we improve data quality by means of data cleaning, outlier processing and missing value processing. In addition, we encode category features for high-frequency category variables and extract deeper information from the limited initial information through feature extraction. A pre-processing scheme of accident duration data is established. Finally, from the perspective of model, sample and parameter diversity, we use XGBoost, LightGBM, CatBoost, stacking and elastic network to build a heterogeneous ensemble learning model to predict the accident duration. The results show that the model not only has good prediction accuracy but can synthesize multiple models to give a comprehensive degree of importance of influencing factors, and the feature importance of the model shows that the time, location, weather and relevant historical statistics of the accident are important to the accident duration.