Yapay Sinir Ağları Destekli Etkinlik Ölçümü: Veri Zarflama Analizi Üzerine Uygulamalar

Abstract

This study reviews the existing methods in the literature of using Artificial Neural Networks (ANN) with Data Envelopment Analysis (DEA), a recent focus of attention. It proposes a new and novel contribution to improving one of the most frequently highlighted vulnerabilities of DEA in the literature: the discriminatory power of DEA. This study aims to explore an absolute and data-driven recipe for the discrimination power is-sues in DEA. Towards this purpose, a new and objective source of information is introduced to the well-established weight-restricted DEA methods. First, founded on the major axioms of DEA, a general production function is modeled via deep learning. Then, inferences on the rela-tive impacts of the inputs on the model predictions are derived-and-translated into ordinal ran-kings by employing machine learning (ML) interpretation algorithms. The two ML interpretation algorithms with different scopes employed in the research are the Input Perturbation Feature Ranking (PFR) and the Shapley Values (SV) algorithms. Moreover, an improvement to the PFR algorithm in removing the inferential variance and adding robustness via substituting a new optimized input re-arrangement with the conventional random shuffling approach. In this research, the use of interpretable ML algorithms in deriving relative importance informa-tion of the DEA criteria is experimented on a single setting and then tested for generalizability via a simulation analysis involving 1000 experiments for each of the applications that employ the PFR or the SV. The experiments concluded that PFR-aided DEA and SV-aided DEA raise matching discriminatory power over the standard DEA models and produce accessible and objective data-driven ordinal rankings of the DEA-input criteria, which computes a complete set of efficiency scores when translated into weight restrictions.