Risky Driver Recognition with Class Imbalance Data and Automated Machine Learning Framework

Area of Science:

• Traffic Safety and Machine Learning
• Computational Driving Analysis

Background:

• Traffic accident prevention requires early identification of high-risk drivers.
• Class-imbalanced driving data poses challenges for standard classification algorithms, often misclassifying minority high-risk driver samples.

Purpose of the Study:

• To propose a novel automated machine learning (AutoML) framework for recognizing risky drivers.
• To simultaneously and automatically optimize sampling techniques, cost-sensitive loss functions, and probability calibration methods to address class imbalance.

Main Methods:

• Utilized Bayesian optimization to tune hyperparameters for sampling ratio and class weight.
• Developed a risky driver recognition model using vehicle trajectory data (longitudinal speed, lateral speed, gap) from 2427 private cars on a German highway.
• Input features derived from discrete Fourier transform coefficients of vehicle dynamics.

Main Results:

• The automated machine learning framework achieved results consistent with manual searching but with significantly improved computational efficiency.
• Identified an optimal combination of Support Vector Machine-based Synthetic Minority Oversampling TEchnique (SVMSMOTE) sampling, cost-sensitive cross-entropy loss, and isotonic regression.
• Demonstrated significant improvements in recognition ability and reduction of predicted probability errors for identifying risky drivers.

Conclusions:

• The proposed AutoML framework effectively handles class-imbalanced driving data for risky driver recognition.
• The optimized combination of SVMSMOTE, cost-sensitive cross-entropy, and isotonic regression enhances model performance.
• This approach offers a computationally efficient solution for improving traffic safety through proactive identification of risky driving behaviors.