A multiobjective view on creating counterfactual explanations for explaining uncertainty in machine learning

Abstract

Counterfactual explanations are a popularly used technique in explainable and interpretable AI. For a given data sample and a trained machine learning model, they provide a new, slightly different data sample together with its new prediction. Recently, counterfactual explanations are also used to explain uncertainty. For that, an optimization problem has to be solved in order to minimize the uncertainty as well as the distance to the original data. This talk gives a multiobjective view on the described problem setting as this provides more insights to the data and its uncertainty.