Publication
The Label Ambiguity Problem in Process Prediction
Peter Pfeiffer; Johannes Lahann; Peter Fettke
In: Cristina Cabanillas; Niels Frederik Garmann-Johnsen; Agnes Koschmider (Hrsg.). Business Process Management Workshops. International Workshop on Artificial Intelligence for Business Process Management (AI4BPM-2022), September 11-15, Münster, Germany, Pages 37-44, ISBN 978-3-031-25383-6, Springer International Publishing, 2023.
Abstract
Predictive process analytics enables proactive situational aw-areness by predicting the future of ongoing process instances. To provide a fair comparison between different approaches developed for process prediction, they have been evaluated on publicly available event logs using the next step prediction task. This paper aims to raise awareness of the label ambiguity problem in the context of process prediction by investigating how uncertainty in the ground truth labels affects next step prediction. Label ambiguity arises from cases in the event log that have different continuation options. We argue that the uncertainty created thereby negatively affects evaluation results. To this end, we present a synthetic example that illustrates the problem of label ambiguity in process prediction and quantify the occurrence of ambiguous ground truth labels in common benchmark datasets. Finally, we discuss implications and present ideas that aim to initiate a discussion on how to deal with label ambiguity in process prediction.
Projekte
SmartVigilance - Regulatorische Compliance durch KI-basierte Umfeldüberwachung in der Medizintechnik