Single-vehicle crash severity outcome prediction and determinant extraction using tree-based and other non-parametric...

Area of Science:

• Traffic Safety Engineering
• Data Science and Machine Learning
• Transportation Research

Background:

• Single-vehicle crashes present significant traffic safety challenges, with a high fatality concentration.
• Existing research predominantly uses econometric models, leaving non-parametric methods for crash severity prediction underexplored.

Purpose of the Study:

• To predict single-vehicle crash severity using various tree-based and non-parametric models.
• To identify key contributing factors influencing crash severity outcomes.
• To compare the performance of different predictive models across various crash severity levels.

Main Methods:

• Utilized Grid-Search for hyperparameter tuning of multiple models.
• Evaluated model performance using accuracy metrics on training, validation, and test sets.
• Performed feature importance analysis to understand contributing factors.

Main Results:

• Models showed comparable performance within the same crash severity level.
• Performance varied across different crash severity datasets, with average training accuracies ranging from 71.76% (PDO) to 99.27% (fatal injury).
• Urban freeways were linked to crash occurrence, while rural freeways correlated with more severe outcomes (fatal/severe injuries).

Conclusions:

• Non-parametric models offer viable alternatives for single-vehicle crash severity prediction.
• Model selection and tuning are critical for optimizing prediction accuracy.
• Future research should explore the impact of temporal variations in contributing features on model performance.