Counterfactual Explanations for Deep Two-Sample Testing
Researchers have developed a new framework for deep two-sample testing that uses counterfactual explanations to identify features driving distributional differences. The method combines a diffusion autoencoder with a pre-trained deep two-sample test model to generate edits that move samples closer to a target distribution. This approach provides interpretable evidence of group differences by showing how specific feature changes affect statistical significance, as demonstrated on synthetic data and MRI cohorts. AI
IMPACT Provides a novel method for interpreting the feature-level drivers of distributional differences detected by deep learning models in statistical tests.