Data Envelopment Learning: An Approach for Feature Selection

This study addresses feature selection in Data Envelopment Analysis (DEA) to mitigate the curse of dimensionality, which causes efficiency score degeneration. We first establish an Exhaustive Combinatorial Strategy as a benchmark to maximize efficiency variance; however, this approach is computationally prohibitive for high-dimensional datasets. To solve this, we propose Data Envelopment Learning (DEL), a hybrid framework integrating Autoencoders for unsupervised feature selection, TreeSHAP for interpretability, and CCR-DEA for evaluation. DEL identifies nonredundant, informative subsets while preserving original features. Experimental results show that DEL achieves near-equivalent performance to the exhaustive search, recovering 99.3% of efficiency variance while reducing computational time by over 104×. Furthermore, the method demonstrates strong ranking stability and superior generalization in cross-validation. These findings indicate that DEL offers a scalable, interpretable, and effective solution for DEA feature selection, enabling its application to complex scenarios that were previously computationally intractable.