Discovering Directly-Follows Complete Petri Nets from Event Data

Abstract

Process mining relies on the ability to discover high-quality process models from event data describing only example behavior. Process discovery is challenging because event data only provide positive examples and process models may serve different purposes (performance analysis, compliance checking, predictive analytics, etc.). This paper focuses on the discovery of accepting Petri nets under the assumption that both the event log and process model are directly-follows complete. Based on novel insights, two new variants ($\alpha$ ${}^{1.1}$ and $\alpha$ ${}^{2.0}$) of the well-known Alpha algorithm ($\alpha$ ${}^{1.0}$) are proposed. These variants overcome some of the limitations of the classical algorithm (e.g., dealing with short-loops and non-unique start and ending activities) and shed light on the boundaries of the “directly-follows completeness” assumption. These insights can be leveraged to create new process discovery algorithms or improve existing ones.