DEA에서 랜덤포레스트를 활용한 단계적 벤치마킹 경로 구성 방법

Abstract

The aim of this paper is to develop a framework for building a stepwise efficiency improvement benchmarking path for decision-making units (DMUs) in data envelopment analysis (DEA) while addressing the curse of dimensionality that often occurs in DEA when there eixsts an excessive number of inputs and outputs relative to number of DMUs. In the framework we first construct all the possible subset combinations of inputs and outputs, and calculate the corresponding DEA models. Secondly, a random forest classification model is trained with the predictor variables being the efficiency results from the DEA models and the target variable being the single performance measure of DMUs. Thirdly, variable importance is computed using the built-in algorithm of random forest, based upon which DEA models of higher significance with respect to the single performance measure can be determined and prioritized. Finally, a stepwise efficiency improvement benchmarking schedule can be established based on the prioritized DEA model results. We illustrate the proposed framework using the financial data for the listed domestic firms in automobile industry.