Traffic Accident Severity Prediction Based on Multi-Model Comparison

Abstract

Traffic accident severity prediction is important for improving traffic safety management and supporting risk assessment and prevention. However, current studies have several limitations with respect to the systematic approach for comparing models and the comprehensiveness of evaluation metrics. This paper aims to systematically assess the overall performance of various machine learning models on a unified experimental framework to predict the severity of accidents in a binary classification task. This study randomly sampled 500,000 records from the US Accidents dataset and used them as the experimental sample. Following data cleaning, missing value handling, feature engineering, and categorical variable encoding, a comparison of four representative models was performed: Logistic Regression (LR), Random Forest (RF), XGBoost and Multi-Layer Perceptron (MLP). With the imbalanced nature of the data, this paper evaluates model performance using accuracy, AUC, ROC curves, and confusion matrices to assess the overall classification performance of different models and their ability to identify the severe accident class. The results show that the tree-based ensemble models Random Forest and XGBoost outperform Logistic Regression and MLP in terms of overall predictive performance, with XGBoost exhibiting better overall performance. Variables that contribute significantly to the model’s predictions include traffic control facilities, time factors, spatial location, and accident-affected distance. The results indicate that machine learning techniques could be useful in traffic accident severity prediction and could be a reference for traffic safety risk assessment. Meanwhile, class imbalance, feature interpretability and model generalizability are challenges that need further investigation and resolution.