RiskFormer: Exploring the temporal associations between multi-type aberrant driving events and crash occurrence

Area of Science:

• Transportation Safety
• Machine Learning
• Data Science

Background:

• Commercial transportation enterprises use driving behavior monitoring for crash risk assessment and intervention.
• Current methods rely on instant event-crash associations, failing to capture time-varying crash risk trends.
• Existing models lack the ability to analyze the temporal coupling influence of multi-type aberrant driving events on crash risk.

Purpose of the Study:

• To explore the temporal associations between multi-type aberrant driving events and crash occurrence.
• To develop a model that accurately depicts the time-varying trend of crash risk.
• To improve proactive interventions in commercial transportation.

Main Methods:

• A contrastive learning method was proposed to analyze single event temporal influence on crash risk, integrating domain knowledge and empirical data.
• A novel Crash Risk Evaluation Transformer (RiskFormer) was developed using a unified encoding method and self-attention mechanism.
• Empirical data from online ride-hailing services were utilized for model training and evaluation.

Main Results:

• Three distinct time-varying crash risk patterns were identified: decay, increasing, and increasing-decay.
• RiskFormer demonstrated a 12.8% improvement in Area Under Curve (AUC) score compared to conventional models.
• The model effectively captures the temporal coupling influence of multi-type events on crash risk.

Conclusions:

• The developed RiskFormer model accurately depicts time-varying crash risk trends by analyzing multi-type events' temporal coupling influence.
• The findings offer a significant advancement in crash risk evaluation for commercial transportation.
• The study highlights the practical utility of advanced machine learning for enhancing road safety and informing proactive interventions.