Applicability of boosting techniques in calibrating safety performance functions for freeways

Area of Science:

• Transportation Engineering
• Data Science
• Machine Learning

Background:

• Safety Performance Functions (SPFs) are vital for traffic safety analysis but often rely on limited regional data.
• Municipalities frequently adapt SPFs from external sources due to data scarcity.
• Boosting algorithms offer a promising approach for improving data analysis and statistical modeling.

Purpose of the Study:

• To evaluate the efficacy of boosting calibration techniques for transferring SPFs internationally using limited regional data.
• To compare the performance of AdaBoost.R2, Two-stage TrAdaBoost.R2, and Gradient Boosting against traditional calibration methods.
• To investigate the impact of training dataset characteristics on the performance of transferred models.

Main Methods:

• Employed AdaBoost.R2, Two-stage TrAdaBoost.R2, and Gradient Boosting algorithms for SPF transfer.
• Utilized a calibration factor method with a negative binomial (NB) regression model as a benchmark.
• Developed two training dataset groups to assess adaptability and the influence of larger datasets.

Main Results:

• Boosting calibration techniques demonstrated superior prediction accuracy compared to the calibrated NB-based model when using limited target region data.
• The amount and distribution of training data were identified as critical factors influencing the proficiency of boosting calibration.
• The study successfully transferred SPFs between U.S. states (Florida, New York) and Chinese cities (Shanghai, Suzhou).

Conclusions:

• Boosting calibration techniques are effective for transferring SPFs in an international context, especially with limited data.
• The performance of transferred models is highly dependent on the quantity and quality of the training dataset.
• These findings support the adoption of advanced machine learning techniques for more accurate traffic safety analysis.