Abstract:
Counterfactual Analysis has been shown to be a powerful tool in the field of Explainable Artificial Intelligence. Most of the contributions in the literature are for Supervised Classification (e.g., whether an individual will be a good payer for a loan or whether a defendant will show no recidivism). Training models with a good trade-off between classification accuracy and transparency is not the only goal. Post-hoc explanations are also sought to give guidance to individuals that have been predicted in the negative class (e.g., those who have been considered as bad payers for a loan or showing recidivism) on how to change their features, with minimum cost, to be predicted in the positive class. The applications of Counterfactual Analysis to other learning tasks are scarce. In this presentation, we illustrate the power of Counterfactual Analysis when applied to the well-known performance measure and benchmarking tool Data Envelopment Analysis (DEA). We define DEA counterfactual instances as alternative combinations of the features that are close to the original features of the firm and lead to desired improvements in the efficiency of the firm. We formulate the problem of finding counterfactual explanations in DEA as a bilevel optimization model. For a rich class of cost functions, reflecting the effort an inefficient firm will need to spend to change to its counterfactual, finding counterfactual explanations boils down to solving Mixed Integer Convex Quadratic Problems with linear constraints. We illustrate our approach using a real-world dataset on banking branches. This work is co-authored with Peter Bogetoft and Jasone Ramírez Ayerbe.