Transfer learning for spatio-temporal transferability of real-time crash prediction models

Area of Science:

• Traffic Safety
• Machine Learning
• Data Science

Background:

• Real-time crash prediction models are crucial for proactive traffic management.
• Spatio-temporal transferability remains a significant challenge for existing models.
• Traffic datasets often exhibit extreme class imbalance, hindering model development.

Purpose of the Study:

• To investigate the spatio-temporal transferability of real-time crash prediction models.
• To address data imbalance issues using Generative Adversarial Networks (GANs) and transfer learning.
• To evaluate the effectiveness of transfer learning in improving model performance across different datasets.

Main Methods:

• Developed a baseline Deep Neural Network (DNN) model using UK motorway data (2017).
• Utilized Wasserstein GAN (WGAN) to generate synthetic crash data for imbalanced datasets.
• Applied transfer learning to adapt the calibrated model to new datasets from different motorways and years.

Main Results:

• Direct transfer of the baseline model was not feasible.
• Transfer learning significantly improved spatio-temporal transferability of crash prediction models.
• Achieved high Area Under Curve (AUC) values (0.69-0.95), with the best model predicting 95% of crashes with a 5% false alarm rate.
• Transferred models performed comparably to or better than standalone models.

Conclusions:

• Transfer learning is effective in enhancing the transferability of real-time crash prediction models, especially under extreme data imbalance.
• The proposed GAN-based transfer learning approach offers a viable solution for developing robust and transferable traffic safety models.
• Findings support traffic engineers in creating more reliable and widely applicable predictive models.