Analysis of the Outcome of the Driving Test for Learner Drivers Based on an Interpretable Machine Learning Framework

Abstract

The driving test is the only way to verify that learner drivers have acquired the competencies stipulated in the national curriculum. Therefore, exploring the key factors that influence the outcome of the driving test is of particular importance in assisting learner drivers to gain solid behind-the-wheel skills. Interpretable machine learning (ML) is employed to analyze the probability of learner drivers’ passing the driving skills test (called the Subject 2 test in China) using a data set comprising personal characteristics, training mode, frequency of driving errors, deducted points, percentage of qualified training times, and score of constructed graphs related to driving behaviors. The data are collected from a driving school in China. A prediction model of the Subject 2 test outcome is constructed by adapting the Light Gradient Boosting Machine (LightGBM) ML method. Furthermore, the SHapley Additive exPlanation (SHAP) is employed to explore the relationships between key influencing factors and the aforementioned outcome. The results indicate that the LightGBM predicts the outcome of the Subject 2 test effectively. The deducted points in the real car training (DP-RC) and the frequency of driving errors in virtual reality (VR) training (FE-VR) have a significant impact on the probability of passing the Subject 2 test.