Business process representation learning

Abstract

Companies generate vast amounts of digital trace data, stored in event logs, when performing business processes. Such data is rich in information, describing what actions have been performed, by whom, at what time, and in what context. Process mining (PM), a data-driven discipline in the field of business process management (BPM), deals with detailed analysis of event data on a process instance basis, aiming to obtain operational insights into how companies’ processes have actually been performed. It has evolved as a major field of research, offering approaches for automatic event log analysis to enhance efficiency or ensure compliance, with growing industrial adoption. As PM applications increasingly require machine learning (ML) techniques, particularly for forward-facing support of business processes, the question becomes how to apply such techniques to event log data, since the characteristics of event logs and the processes they describe differ from common data modalities such as images and text. Given the complexity of event log data, applying ML techniques poses significant challenges, requiring adaptations of existing or the development of customized ML techniques to account for the characteristics of event logs. Representation learning, a research field in ML, deals with learning representations from data that make solving downstream tasks more effective and efficient by overcoming the laborious and expensive task of manual feature engineering. This thesis introduces the problem of business process representation learning, i.e., learning representations from event logs for solving BPM tasks like process prediction, anomaly detection, or task abstraction. By developing, analyzing, and applying various process representation models (RPMs), this thesis contributes to three research areas. First, on how to design PRMs in terms of architecture and training procedures to learn representations that follow the general priors of representation learning. Second, on the assessment of how well the models have learned characteristics of event log data and the underlying processes. Finally, on the application side, for which BPM tasks PRMs can be used, and how well they perform. The results demonstrate that PRMs are capable of learning accurate, context-specific representations from different concepts within event logs. These representations adhere to the principles of representation learning and can be utilized to solve real-world BPM tasks efficiently. Thereby, the results advance the fields of PM and ML, and especially their interplay.