Hybrid machine learning-based model for evaluating the performance of agile-sustainable supply chains in the context of industry 4.0: a case study

In today’s world, businesses and, in general, supply chains have undergone extensive transformations, and relying solely on traditional metrics such as cost and quality cannot provide a comprehensive and complete evaluation of companies active in various sections of supply chains. One of the main concerns of supply chain managers is to create an integrated and comprehensive structure for evaluating the performance of active branches. In this context, this study presents a structure that, by simultaneously considering agility and sustainability metrics within the context of the industry 4.0, which has brought about fundamental changes in the supply chain environment in recent years, aims to evaluate the active branches in the dairy product supply chain. On the other hand, the increase in the volume of data produced in the supply chain environment and the development of the applications of machine learning algorithms in various fields, which offer better applications compared to intuitive approaches, have led this study to use hybrid data-driven approaches, which are a combination of expert-based methods and documented organizational data, to evaluate the performance of supply chain branches. Therefore, this study is innovative in terms of the evaluation metrics and the data-driven approach developed. In the first step, evaluation metrics appropriate to the dimensions of agility, sustainability, Industry 4.0, and general metrics were identified, and then the fuzzy best-worth method (FBWM) approach was used to weight the metrics. According to the findings, data-driven, marketing, overhead costs, delivery timeframe, and product quality were selected as the most important metrics. Subsequently, using the developed artificial neural network algorithm, which calculates the input weights of the metrics using the FBWM method, a model for evaluating the supply chain was presented, and the findings show that the developed approach performs better than other algorithms on the problem data with more than 92% accuracy.