Explainable AI through Variable Importance Tests – VITE

The use of Artificial Intelligence (AI) techniques in natural sciences have become pervasive, owing to their ability to perform prediction of a given outcome given some potentially complex input data. However, for the sake of scientific knowledge, AI cannot be used as a ready-made recipe, but should lead to scientific explanations, requiring the use of Explainable AI (xAI). While xAI can correspond to relatively simple models when few variables are involved, it becomes a hard task in a high dimensional setting.
VITE thus aims at building reliable tools with precise statistical guarantees to understand the impact of a given variable used to predict some outcome of interest. The approach should be model agnostic. For this, we propose to extend conditional permutation schemes proposed recently. We want to study in depth the theoretical guarantees that follow from such procedures. Second, to better account for the complex dependencies that exist between variables, we want to build tests that captures those dependencies, by generalizing the use of sampling methods.

Such sampling methods can be viewed as a special case of the now popular generative models in AI.
Finally, we would like to make such procedures available to users. For this, we will edit a software package using an API compatible with scikit-learn.
This will include examples showcasing the use of variable importance in several realistic examples coming from brain imaging genetics.