Leveraging Generative Adversarial Networks for Improved Red-Light Running Crash Severity Modeling

Red-light running (RLR) is a major safety concern at signalized intersections, driven by both intentional and unintentional factors. This complexity makes it difficult to develop universal countermeasures and underscores the need for a deeper understanding of factors contributing to severe RLR crashes. However, crash severity modeling is frequently hampered by class imbalance in crash datasets, which can introduce bias and undermine predictive reliability. Recent advances in generative artificial intelligence, particularly generative adversarial networks (GANs), offer a promising solution by synthesizing realistic minority-class samples to address class imbalance. This study evaluates the effectiveness of conditional tabular GANs (CTGANs) against traditional resampling techniques—random over-sampling (ROS), synthetic minority over-sampling technique, adaptive synthetic sampling, and tabular variational autoencoder—using a dataset of 20,430 RLR crashes in Alabama, US, from 2015 to 2024. Results show that CTGAN, when combined with a multi-layer perceptron (MLP) model, achieves the best overall predictive performance among all evaluated combinations, with a geometric mean (G-mean) of 0.55 and an accuracy of 0.69 on the test set, outperforming the next best alternatives, CTGAN + support vector classifier (accuracy 0.67, G-mean 0.54) and ROS + MLP (accuracy 0.66, G-mean 0.54). Model interpretability using SHAP analysis revealed that the most important factors in predicting injury-related RLR crashes were the involvement of vulnerable road users (i.e., pedestrians, bicyclists, and motorcyclists), multi-vehicle crashes, heavy-vehicle involvement, nighttime conditions, rural areas, and intersections with five or more approaches. These findings provide valuable insights for developing policy recommendations and effective countermeasures to reduce injury-related RLR crashes.