This study presents a novel framework that integrates Data Envelopment Analysis (DEA) with machine learning (ML) classification techniques to improve robustness, interpretability and relevance of efficiency assessments. The method reinterprets DEA as a probabilistic classification task, using ML to estimate inefficiency scores and the probability of a decision-making unit (DMU) being efficient. This probabilistic approach enables uncertainty-aware benchmarking and decision support. Explainable AI (XAI) techniques, particularly Shapley Additive exPlanations (SHAP), are used to identify the most influential input-output variables and guide resource reallocation strategies. This allows the definition of data-driven directional vectors for counterfactual analysis. Peer selection is dynamically adjusted based on proximity and efficiency probability thresholds, enhancing benchmarking relevance. The methodology is algorithm-agnostic and can be implemented with alternative ML classifiers. Applied to 917 firms in the Spanish food industry, the framework yields interesting results. In the study, neural networks achieved the best predictive performance, with a balanced accuracy of 87%, compared with 81% for SVM, 80% for random forests, and 69% for logistic regression. The probabilistic approach also improves discrimination relative to deterministic DEA: while DEA identifies 29 firms as efficient, only 25 remain efficient under a probability threshold of 0.85. At the industry level, reaching this threshold requires average input reductions of 27% and output increases of 59%. Finally, firms are ranked into three groups: those already above the target probability threshold, those capable of reaching it through feasible adjustments, and those constrained by a lower maximum attainable probability.
González-Moyano et al. (Fri,) studied this question.