Standard Neural Network Outperforms Hybrid Models as Surrogate for Big-Data Efficiency Analysis

Researchers report that a plain artificial neural network delivered stronger generalization and more reliable classification than two hybrid variants when used to classify decision-making units in data envelopment analysis at big-data scale. The study, published in Neural Computing and Applications by A. Fakharzadeh J. and V. Amiri, tackles a concrete bottleneck: traditional data envelopment analysis, or DEA, faces computational complexity and growing calculation time when datasets become very large, particularly when the number of decision-making units is huge or the input-output space is high-dimensional. The authors integrated machine learning techniques into the DEA framework to classify these units efficiently, testing three approaches: a standard artificial neural network, a genetic algorithm neural network in which evolutionary search is coupled to network training, and a particle swarm optimization neural network that adds swarm-based search to the training process.

The proposed models were evaluated on two datasets, and their performance was assessed with several complementary measures: computational time, mean squared error, accuracy metrics, and misclassification rate. The study also investigated how different network configurations affect the results, including the choice of activation function and the number of hidden neurons, mapping how architecture choices shape the behavior of the surrogate models. Across these evaluations, the standard artificial neural network showed stronger generalization and more reliable classification performance than the hybrid GANN and PSO-ANN models.

According to the authors, the plain ANN achieved a better balance between computational efficiency and frontier preservation, meaning it preserved the efficiency frontier that DEA is meant to characterize while reducing the heavy calculation burden of running the full analysis at scale. The added machinery of genetic algorithms and particle swarm optimization did not translate into better classification of efficient versus inefficient units in this setting. The authors describe the standard ANN as a promising surrogate for large-scale DEA evaluation, a finding that stands out precisely because the simpler network, rather than the more elaborate hybrids, proved the stronger candidate in the reported tests.

The comparison rests on two datasets, and the reported advantages of the standard ANN apply to the configurations and evaluation measures examined in the study, including computational time, mean squared error, accuracy metrics, and misclassification rate. Within that scope, the results indicate that coupling a standard neural network to the DEA framework can screen very large panels of decision-making units more efficiently than classical DEA alone, while the hybrid GANN and PSO-ANN models offered no classification advantage in the reported evaluations.

Subject of Research: Detecting efficiency based on machine learning and neural networks approaches in DEA involving big data

Article Title: Detecting efficiency based on machine learning and neural networks approaches in DEA involving big data

Article References: Fakharzadeh J., A., & Amiri, V. (2026). Detecting efficiency based on machine learning and neural networks approaches in DEA involving big data. Neural Computing and Applications, 38(20), Article 843. https://doi.org/10.1007/s00521-026-12600-3

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12600-3

Keywords: Detecting, efficiency, based, machine, learning, neural, networks, approaches, involving, data, scientific research