Prof. Dr. h.c. mult. Hasso Plattner

Machine Learning for Sales Order Fulfillment

Large enterprises and their information systems produce and collect large amounts of data related to different areas, e.g., manufacturing, finance or human resources. This data can be used to complete tasks more efficiently, automate tasks that are currently executed manually, and also generate insights in order to solve certain challenges.

Project Members from left to right: Nils Thamm, Jonathan Schneider, Lukas Ehrig, Christian Flach, Philipp Bode, Hendrik Rätz, and Hendrik Folkerts (Foto: HPI/Herschelmann)

Nowadays, machine learning techniques are utilized in many fields and use cases. In cooperation with the SAP Innovation Center Network, this bachelor project investigated opportunities to apply machine learning techniques to the problem of order delay prediction. The recent development and in-practice application of in-memory database technology (e.g. SAP HANA) enabled the efficient execution of these techniques on large enterprise datasets.

Edison: Predictive Analytics for Order Fulfillment

Delays and problems during sales order fulfillment processes are caused by a multitude of factors. Oftentimes issues causing the delay only become obvious when the process is already interrupted. To improve the handling of these issues, one needs to analyze and extract the factors commonly responsible for delays.

Landingpage of the Edison application that gives an overview on the estimated delays of sales orders caused by predicted issues.

Enterprise systems contain often data of millions of historic sales orders. These can be utilized to draw conclusions for current and upcoming orders. Based on real world data of historic sales order fulfillment processes, we apply decision trees that estimate which issues may interrupt the process for a currently open order. This flexible machine learning technique enables to display details of the most pressing sales orders on the one hand, and furthermore to review the reasons why a particular sales order is considered to become delayed on the other hand. Therefore, the user receives the relevant information about critical indicators of an open sales order right at the beginning of the fulfillment process. 

In order to proactively use the received information, a partially observable Markov decision process (POMDP) model is applied to propose the best long-term action. Therefore, based on the previously derived risks for issues of an open sales order, and the information about all action’s costs and consequences the user receives the best action to mitigate a delayed sales order. 

Additional Details


Johannes HuegleJan Kossmann, Dr. Matthias Uflacker