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

Area of Science:

• Traffic Safety Engineering
• Data Science
• Machine Learning

Background:

• Crash frequency modeling is crucial for road safety but hindered by rare crash events and imbalanced datasets.
• Existing methods struggle with heterogeneous multi-type crash data and strict distributional assumptions.
• Deep generative models often distort data or fail with diverse variable types.

Purpose of the Study:

• To propose TabSyn, a hybrid Variational Autoencoder (VAE)-Diffusion model for augmenting imbalanced multi-type crash data.
• To preserve inter-variable correlations and underlying data structure in synthetic crash data.
• To improve crash frequency prediction and identification of high-risk road segments.

Main Methods:

• TabSyn utilizes a tokenizer-based embedding and VAE encoding to map multi-type crash data into a continuous latent space.
• The model was compared against state-of-the-art generative models (CTGAN, TVAE, GReaT, StaSy) for synthetic data quality.
• TabSyn was integrated with XGBoost for crash frequency prediction, evaluated against ZIP and GAM-Poisson models.

Main Results:

• TabSyn achieved superior synthetic data distributional and structural fidelity compared to benchmark generative models.
• The model demonstrated the lowest prediction error in crash frequency modeling.
• TabSyn excelled in identifying Top-K high-risk road segments.

Conclusions:

• TabSyn offers an effective solution for imbalanced multi-type crash data augmentation in traffic safety.
• The proposed method enhances the accuracy of crash frequency prediction models.
• This research provides valuable insights for improving road safety management through advanced data augmentation techniques.