A F-R Algorithm-based Machine Learning Model for Traffic Accident Severity Prediction

Abstract

Prediction of severity of traffic accidents has been a crucial subject area within intelligent transportation systems, as correct estimation plays a fundamental role in decision-making processes in emergencies. Traditional approaches to accident severity prediction using cumulative logistic regression models often make assumptions about linear interrelationships between the input data, and thus do not work well with nonlinear interactions typical of traffic data. This paper proposes a novel approach to optimizing intelligent transportation models using the hybrid Foraging Habitat Selection Particle Swarm Optimization–Random Forest (FHSPSO-RF) technique. In particular, a new FHSPSO algorithm can adaptively optimize several parameters of the Random Forest classifier, namely the number of trees, maximum depth, and minimal samples per split, which greatly contributes to its effectiveness. The experiment results confirm that the optimized algorithm yields superior results in terms of accuracy and F1-score compared with conventional Support Vector Machine, K-Nearest Neighbor, Logistic Regression, and Random Forest algorithms. Besides, SHAP analysis helps to identify important contributive features in the process of accident severity prediction, including number of injuries, number of vehicles, lighting, and weather conditions.