Using contextual data to predict risky driving events: A novel methodology from explainable artificial intelligence

Area of Science:

• Data Science
• Transportation Safety
• Insurance Telematics

Background:

• Usage-based insurance leverages driving behavior data for premium adjustment.
• Telematics data offers insights into driving contexts (road type, weather, traffic).
• Driving contexts significantly influence accident exposure and risk.

Purpose of the Study:

• Investigate the relationship between driving context combinations and driving risk.
• Identify and rank contextual factors predicting near-misses, speeding, and distraction events.
• Provide insights for road safety stakeholders and insurers.

Main Methods:

• Utilized a naturalistic driving dataset (77,859 km).
• Employed XGBoost and Random Forests for predictive modeling.
• Applied Shapley Additive Explanations to identify and rank feature importance.

Main Results:

• Driving context is a significant predictor of driving risk.
• Key predictors include speed limit, temperature, wind speed, traffic, and road slope.
• Low speed limits increase speeding events; low temperatures decrease harsh maneuvers; precipitation increases harsh maneuvers and distractions.

Conclusions:

• Driving context analysis enhances road safety and insurance risk assessment.
• Specific contextual factors have predictable impacts on various risky driving events.
• Methodology supports data-driven strategies for mitigating road accidents.