Ensemble based Explainable Demand Forecasting for Industrial Electronic Products

This paper presents a novel machine learning (ML) architecture to address the issue of predicting the demand for a business-to-business (B2B) product used in generic industrial applications that exhibits dramatic fluctuation over time, through ensemble and optimization-based approach. In this scenario a set of related indices are used as features to train the ML models. It is observed that even the predictions generated through a single model varies dramatically across hyperparameters. In the proposed approach, large number of predictions are combined through a novel generic optimization-based layer to generate the final prediction, in a hierarchical manner. Since many popular commonly used loss functions are convex in nature this layer can be implemented using matured embedded modeling language with open-source solvers. To improve the adoption of the proposed forecast system we provide explainability with respect to the related indices that are well understood by the business managers. Towards that end we leverage an advanced method GAMI-Net based on neural network to compute the feature contributions and provide useful insights to the user that offers intrinsic explanation. We compare the performance between the GAMI-Net and SHAP which is more popularly used in literature. Through case studies with real dataset for an industrial B2B product we highlight the advantages of GAMI-Net over SHAP. We perform exhaustive studies to evaluate the performance of the proposed architecture with respect to three existing state-of-the-art methods for forecasting. The results demonstrate that the proposed architecture significantly improves the robustness and accuracy of the forecast and provides better explainability.