Explainable Ensemble of Deep Learning Models for Improved and Interpretable Product Backorder Prediction

Abstract

Accurate Backorder prediction plays a critical role in improving decision-making and enhancing the accuracy of the supply chain risk assessment. The application of deep learning (DL) models to inventory data provides a powerful framework for obtaining high predictive performance for this task. Nevertheless, DL models often operate as lack of interpretability, making their predictions difficult to interpret and trust in inventory management applications. This paper explores the design of backorder prediction models that balance accuracy and explainability to ensure transparency and trust among supply chain decision makers. We propose a two-phase explainable ensemble of deep learning models (EEDL) for product backorder prediction. The proposed backorder prediction task is formulated as a binary classification problem (backorder vs. no-backorder). In the first phase, we propose an ensemble of deep learning models utilizing four deep classifiers, namely MLP, DNN, CNN and LSTM, selected for their complementary capabilities in capturing nonlinear relationships, temporal dependencies, and complex feature interactions within backorder data. In the second phase, the explainability of EEDL outputs is enhanced using two eXplainable Artificial Intelligence (XAI) methods: LIME for local explanations and SHAP for both local and global explanations. Unlike existing approaches in backorder prediction, this work introduces a structured heterogeneous deep learning ensemble, combined with a novel hybrid interpretability mechanism designed for ensemble learning rather than a single deep learning model. Experiments are conducted on a real-world dataset containing 1,929,936 records with 23 features, including lead time, historical sales and inventory levels. Experimental results show that the proposed EEDL model achieves high predictive performance, with an accuracy of 97.61% on the real-world dataset. Furthermore, comparative analysis demonstrates that EEDL outperforms several existing backorder prediction models, confirming its effectiveness in improving both prediction accuracy and explainability in supply chain management.