Machine learning-enhanced Data Envelopment Analysis via multi-objective variable selection for benchmarking combined electricity distribution performance

The Data Envelopment Analysis (DEA) model can be interpreted as a machine learning-based Sign-constrained case of Convex Nonparametric Least-Squares (SCNLS) regression. However, the selection of input variables in SCNLS-based DEA for benchmarking electricity distribution performance is only for the single-output setting. This paper proposes a Multi-task Least Absolute Shrinkage and Selection Operator (M-LASSO)-enhanced DEA framework that evaluates combined performance scores of distributors in an end-to-end manner. A benchmarking criterion associated with M-LASSO-enhanced DEA, namely performance satisfaction measure, is defined to facilitate the multi-objective variable selection in M-LASSO-enhanced DEA. For implementation, we develop the Corrected M-LASSO (CM-LASSO) method as a two-stage solution approach for the M-LASSO-enhanced DEA model. Using Monte Carlo simulation data, the comparison among SCNLS-based, LASSO-enhanced, and M-LASSO-enhanced DEA models indicates the effectiveness of multi-objective variable selection. Using a real-world dataset from 1993 to 2021, the performance scores of distributors in China demonstrate a trend towards increasingly converging composite performance. This trend, which contains economic efficiency, supply reliability, and environmental sustainability, supports further deregulation in the context of achieving emission peak targets and advancing towards carbon neutrality.