Decentralized Predictive Enterprise Resource Planning Framework on Private Blockchain Networks Using Neural Networks

Abstract

Theoretically, cross-department predictive modeling can improve the operational efficiency of an enterprise, particularly on enterprise resource planning. For example, a model that predicts the volume of purchase goods will be more generalizable if the predication is based on the data from multiple departments. Most existing cross-department predictive models rely on a centralized technology, in which security and robustness are ignored, including unreliable single-point or malicious modification of records. Therefore, our works propose a decentralized framework to combine Blockchain technology with exited model so as to apply in predictive enterprise resource planning. In detail, model parameter estimation will be trained by without revealing any other information, which means only model-related data are exchanged across departments. In order to apply transaction metadata to disseminate models, we introduce neural networks combine with a private Blockchain network. In addition, we design an algorithm to train the neural networks that combine the loss function from each local model to achieve the smallest global level validation loss. Finally, we implement the experiments to prove the effectiveness of our framework by applying it to multi typical tasks in enterprise resource planning. Experimental results reveal the advantages of this framework on both tasks.