Improving the transferability of the crash prediction model using the TrAdaBoost.R2 algorithm.

Area of Science:

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

Background:

• Crash prediction models are vital for traffic safety but face challenges with limited or low-quality data in some jurisdictions.
• Existing calibration methods for transferring crash prediction models, such as aggregate and disaggregate approaches, have limitations in accuracy and resource consumption.
• Transfer learning offers a promising alternative for calibrating models by leveraging knowledge from existing data domains.

Purpose of the Study:

• To evaluate the effectiveness of an instance-based transfer learning technique, TrAdaBoost.R2, for site-based crash prediction model transfer.
• To compare the predictive performance of TrAdaBoost.R2 against the conventional calibration factor method and AdaBoost.R2.
• To assess the adaptability of TrAdaBoost.R2 to small sample sizes in the target data domain.

Main Methods:

• An instance-based transfer learning algorithm, TrAdaBoost.R2, was employed for model transfer.
• TrAdaBoost.R2 was compared with AdaBoost.R2 using pooled data to assess knowledge extraction from an outdated source domain.
• The negative binomial model's calibration factor method was used as a benchmark for comparison, with mean square error evaluating prediction accuracy.

Main Results:

• Models developed using TrAdaBoost.R2 demonstrated superior prediction accuracy compared to the conventional calibration method.
• TrAdaBoost.R2 showed better predictive performance and adaptability to small sample sizes, making it suitable for site-based transfers.
• The study utilized data from Shanghai and Guangzhou, China, as mutual source and target domains to validate the transfer learning approach.

Conclusions:

• The TrAdaBoost.R2 transfer learning technique is recommended for developing accurate and adaptable crash prediction models, particularly when dealing with limited data.
• Separate crash prediction models for peak and off-peak hours are proposed to further enhance prediction accuracy.
• Transfer learning offers a viable solution to overcome data limitations in traffic safety analysis.