A Machine Learning Approach for Solving Big-Data Network DEA Models

This study proposes a machine learning-based approximation framework to alleviate the substantial computational burden of two-stage network Data Envelopment Analysis (DEA) in large-scale applications. Although network DEA provides a more realistic representation of multi-stage production processes than conventional single-stage models, its computational cost grows rapidly as the number of decision-making units (DMUs) increases, limiting its applicability in big-data environments. To address this challenge, the proposed framework evaluates only a strategically selected subset of DMUs using an exact two-stage network DEA model and subsequently trains supervised learning models to approximate stage-wise and overall efficiency scores for the remaining units. Both regression and classification approaches are considered to capture efficiency levels and frontier membership. In addition, a frontier projection-based data augmentation scheme is introduced to mitigate data sparsity near the efficient frontier and improve model generalization. Extensive simulation experiments demonstrate that the proposed approach reduces computational time by up to 90% while preserving high predictive performance. In particular, tree-based models achieve coefficients of determination exceeding 0.95, and efficiency classification yields F1 scores above 0.98. These findings indicate that the proposed framework enables scalable and near-real-time efficiency evaluation in complex network production systems, thereby extending the practical applicability of network DEA to large-scale and data-intensive decision-making contexts.