A meta-learning approach to improving transferability for freeway traffic crash risk prediction

Crash risk prediction plays a vital role in preventing freeway traffic accidents. Due to the limited availability of crash data in some freeway sections, model transferability of crash risk prediction has become an essential topic in traffic safety research. However, only limited research has been conducted on transferability improvement and applications of existing models in large-scale freeway networks. This study presents a Model-Agnostic-Meta-Learning (MAML) based framework to improve transferability and robustness, which is applicable to any crash risk prediction model trained with gradient descent. The proposed framework is trained and tested using the freeway crash records from the Caltrans Performance Measurement System (PeMS) in 2017 and 2022. The results show that the proposed framework effectively avoids over-fitting and performs better on spatial and temporal transferability in both single-task learning, and multi-task learning. Three benchmark models are developed to compare the results with the proposed framework, demonstrating that the MAML-based method leads to state-of-the-art performance. The distributions of multi-learning results are also plotted to understand the effects of the framework and reveal how the proposed framework improves the model performance. The findings indicate the promising performance of using the proposed meta-learning approach to enhance current models for crash prediction and freeway safety management.