Séminaire hybrique sur Zoom et dans la salle de séminaire du GERAD.

Counterfactual explanations describe how to modify a feature vector to flip the outcome of a trained classifier. Several methods have been proposed to generate counterfactual explanations, but their robustness when the classifier is re-trained has not been studied so far. Our goal is to obtain counterfactual explanations for random forests that are robust to algorithmic uncertainty. We study the link between the robustness of ensemble models and base learners and formulate a chance-constrained optimization model. We provide statistical guarantees for random forests of stumps and develop a practical method with good performance. We show that existing naive and plausibility-based methods provide surprisingly low robustness. Our method achieves the best trade-off between robustness and the distance of counterfactual explanations to the initial observation.

Séminaire conjoint avec Dounia Lakhmiri.