Reproduce “Generating collective counterfactual explanations in score-based classification via mathematical optimization” on OpenResearch

About this paper

Emilio Carrizosa, Jasone Ramírez-Ayerbe, Dolores Romero Morales

This work introduces a mathematical optimization framework to generate collective counterfactual explanations for groups of instances in score-based classification models, offering insights into group-level interventions rather than just individual changes. It successfully demonstrates the identification of globally critical features and outliers across various real-world datasets using logistic regression and additive tree models.

Key insights

• Mathematical optimization provides a rigorous framework for generating optimal collective counterfactuals, ensuring minimal changes under specified constraints.
• The global sparsity term effectively reveals a small, universally critical set of features whose collective modification influences the classification for many individuals in a group.
• The framework can identify outliers within a group, highlighting instances that are too costly to perturb collectively and may require individual attention.

Summary by alphaXiv.