Comparing Architectures of Neural Networks for an Integration in Enterprise Systems: A Retail Case Study
Marek Hütsch · Procedia Computer Science · 2021
While standard software has largely established itself on the market for business software, the unified integration of machine learning algorithms in standard software has received little attention so far. Business standard software serving as main component of an enterprise system is traditionally divided into software modules, such as CRM and SRM. This paper examines the architecture of neural networks (NN) with the aim of integrating them into business standard software. For this purpose, two optimization goals are chosen, the demand forecast and the delivery forecast. The delivery forecast has the aim to make recommendations for article replenishment deliveries for retail stores. These two optimization goals are traditionally implemented in different standard software modules, CRM and SRM, which differ in the nature of their task: Analytical (demand forecast) and transactional (article deliver forecast). Two different architectures of NNs are investigated. It turns out that a modularization of the NN with respect to their forecasting goal combined with a meaningful linking of the networks leads to a significant forecast precision improvement.