One-Class SVM for Estimating Production Technology: An Investigation Based on DEA Microeconomic Axioms

Abstract

A significant methodological challenge in Data Envelopment Analysis (DEA) is its inherent vulnerability in overfitting. This occurs because the DEA frontier is constructed empirically just from observed data points, causing the frontier to "wrap" too tightly around the dataset. Consequently, the corresponding evaluation model may suffer from random noise, measurement errors, or extreme values for genuine technological relationships. This over-adaptation hence leads to optimistically biased efficiency scores, reduces the model's discriminatory power in distinguishing between units. Moragues et al. (2023) developed an approach for constructing a new production possibility set using One-Class Support Vector Machine (OC-SVM) model which leads to a more robust and less sample-sensitive efficiency frontier. The aim of this paper is to investigate the main properties of the production technology formulated by Moragues et al. (2023) from a microeconomic points of view, and to demonstrate its advantages in overcoming some limitations of the conventional DEA approach.