빅데이터 DEA를 위한 기계학습의 적용

Abstract

Data envelopment analysis (DEA) is a tool for identifying best-practices when multiple performance metrics or measures are present for decision-making units (DMUs). As big data issue becomes an important area of supply chain and operations management, DEA is evolving into a data-oriented data science tool for benchmarking, performance evaluation, composite index construction and others. As the number of DMUs increases, the running-time to solve the standard DEA model sharply rises. Such situations are appearing more frequently in the era of big data. This issue could be an important challenge particularly when real-time data stream in at extremely high rates and the DEA analysis needs to be performed very quickly. Therefore, there exist practical needs for developing an efficient way of solving large-scale DEA problems. In this paper, we propose a practical approach for speeding up the DEA efficiency estimation process based on machine learning. In this approach, a sample of DMUs is selected from the population as a training data set, based on which a machine is trained to predict the efficiency scores of unselected or newly streamed-in DMUs. We also suggest a data augmentation technique to enhance the learning process under severe data class imbalance. The superior performance of the proposed approach over the conventional one in terms of efficiency prediction power as well as model computation time is shown through a series of computational experiments using randomly generated data.