VAE-Inf: A generative framework for imbalanced classification

A new two-stage framework called VAE-Inf integrates variational autoencoders with hypothesis testing to address imbalanced classification when minority samples are extremely scarce. In the first stage, a VAE is trained exclusively on majority-class data to model the reference distribution, and latent posteriors are aggregated via a Wasserstein barycenter into a global Gaussian reference. The second stage transforms this representation for statistical inference, enabling reliable error control and stable decision boundaries. The approach bridges generative modeling and discriminative classification, offering a statistically interpretable solution for extreme class imbalance.