Novel Analytics and Data Generation Methods Towards Road Crash Data Scarcity

Road crash analysis suffers from critical data scarcity across various data types: imbalanced severity classes, zero-inflated frequency counts, and scarce secondary crash records. This thesis develops a unified generative AI framework to address these diverse data challenges. For crash severity, CTGAN-RU synthesizes discrete & continuous data; for crash frequency, a VAE-Diffusion model generates multi-type data; for secondary crashes, an LSTM-GAN-Transformer produces integrated dynamic & static features. Validated on multiple years of highway data, the framework significantly outperforms existing models across all prediction tasks, providing robust solutions for various crash data analysis scenarios under data scarcity conditions.