Robust Directional Counterfactual Explanations for Black-Box Models

J. Martín-Chávez, E. Carrizosa, C. Molero-Río

We propose a framework for directional counterfactual explanations to interpret black-box models.
This approach identifies an optimal direction of change that maximizes robust predictive improvement.
By ranking directions through a lower-tail CVaR functional, the method focuses on reliable gains under unfavorable realizations.
We establish regularity guarantees, derive exact evaluation schemes for common model classes, and propose vMF--VNS, a derivative-free stochastic optimizer.

Keywords: Machine Learning Optimization Models and Methods Robust Optimization

E. Carrizosa, J. C. Castro Gómez, V. Guerrero

M. Alonso-Pena, E. Carrizosa, A. Navas-Orozco

E. Carrizosa, R. DE LEONE, M. MAGAGNINI

C. Molero-Río, C. D'Ambrosio