Distribution Shift

We answer the question: ‘‘What is a distribution shift explanation?’’ and introduce a novel framework for explaining distribution shifts via transportation maps between a source and target distribution which are either inherently interpretable or interpreted using post-hoc interpretability methods.

Focusing on distributions shifts pertaining to images, we use interpretable transport maps between the latent image spaces of a source and a target distribution to explain how to align the source to the target distribution.

We formalize the problem of feature shift, and introduce a method for fast and simultaneous detection of domain shifts and localizing the shift to specific feature(s).