A data analytics-driven approach to backorder prediction using federated machine learning in industrial supply chains

In the modern era, businesses and production industries are increasingly adopting innovative technologies and data-driven approaches to optimize their supply chain processes. In this regard, promptly handling backorders has a significant impact on enhancing the company’s supply chain efficiency and inventory management. Generally, backorders occur when a product is no longer available or completely out of stock, and a client makes an order for subsequent manufacturing and distribution. The major challenges in backorder prediction systems are that data from the supply chain frequently includes sensitive information regarding inventories, suppliers, and client orders. In addition, the supply chain is often distributed among different warehouses, stores, distributors, retailers, collaborators, and stakeholders. These parties are often reluctant to share important data because of privacy and security concerns. To address this problem, this study proposes a secure, privacy-aware backorder prediction system using federated learning at the intersection of engineering and supply chain management. While existing methods predict backorders, they usually lack privacy, security, and decentralized data handling. Hence, to provide more secure early backorder prediction, this study employs single- and multi-layer neural networks trained in a federated setup. The class imbalance challenges of the dataset are also addressed using SMOTE and Near-Miss Sampling methods. Moreover, the proposed model’s explainability is also studied using LIME (Local Interpretable Model-agnostic Explanations) and SHAP (Shapley Additive explanations). The results show that the proposed models show good results over other centralized machine learning algorithms as well as state-of-the-art methods, with a highest accuracy of about 90.3%.