Explainable Remaining Time Prediction for Business Processes

Predictive process monitoring is a central practice in business process management that allows for the timely identification of errors, bottlenecks or deviations in process behavior, i.e. actionable information. Established predictive monitoring methods use event data extracted from Workflow Management Systems in which the case is perceived as the primary element of the process and as a result are approached under the wrong assumption that cases behave in isolation. Because of this intra-case perspective, current predictive monitoring techniques are not able to detect inter-case dynamics that do emerge in organizational processes such as batching. In this thesis we address the need for the awareness of inter-case dynamics in the development life cycle of predictive monitoring approaches specifically aimed at remaining time prediction. We propose a first set of techniques that together introduce inter-case dynamics in the remaining time prediction life cycle. We present techniques for identifying sub-sets of cases with high prediction errors that are subject to these inter-case dynamics and we present techniques that use these insights to create inter-case features and derive an inter-case evaluation specifically for inter-case dynamics caused by batching. In an experimental setup we show that by applying this set of techniques, we are able to improve prediction performance and a better evaluation thereof, leading to more adequate and better explainable models and results.Master thesis