Colloque / Séminaire

Abstract:

Machine learning models are increasingly used to make crucial decisions such as credit insurance approval or diagnose patients. 

Post-hoc interpretability approaches have been proven to be powerful tools to generate explanations for the predictions made by a trained black-box model. This presentation proposes two post-hoc approaches to detect the closest decision border of a classifier and generate local explanations for a given prediction. The first one tries to generate counterfactual explanations when no information is available about the classifier nor any data. On the other hand, the second one tries to approximate the decision border of the classifier with a surrogate model to generate more stable explanations. In particular, the notion of local explanation is studied through these two approaches.