Traffic crash data augmentation with multi-type variables using hybrid VAE-Diffusion generative neural networks for...

Crash frequency modeling aims to analyze influential factors of crashes to enhance road safety. However, as crashes are inherently rare events, excessive zero observations in crash datasets undermine crash frequency models' ability to identify high-risk road segments. Existing statistical models are often limited by strict distributional assumptions, while resampling and deep generative methods often distort data representation or struggle with multi-type crash data (count, ordinal, nominal, and real-valued) since these heterogeneous variables follow different statistical distributions and require distinct encoding strategies. This study proposes TabSyn, a hybrid VAE-Diffusion model that transforms multi-type crash data into a continuous latent space through a tokenizer-based embedding and VAE encoding, enabling the model to preserve the inter-variable correlations and underlying data structure. To validate its effectiveness, TabSyn is compared with state-of-the-art generative models (CTGAN, TVAE, GReaT, and StaSy) in synthetic data quality, and integrated with XGBoost for crash frequency prediction against two statistical models (ZIP, GAM-Poisson). Results demonstrate that TabSyn outperforms benchmark methods by achieving the best synthetic data distributional and structural fidelity, and the lowest prediction error, particularly for Top-K high-risk segment identification. This study offers valuable insights for improving crash frequency modeling and traffic safety management through imbalanced data augmentation of multi-type crash data.