Diffusion Autoencoder Method Adds Counterfactual Explanations to Deep Two-Sample Tests

A counterfactual explanation framework for deep two-sample testing uses a diffusion autoencoder paired with a pretrained test model, optimizing an MMD objective in representation space to generate sample-level edits that explain why a null hypothesis is rejected. The approach targets high-dimensional structured data like images where classical kernel tests underperform.