Publikation
Synthetic Log Generation Under Control-Flow Conditions Using Autoregressive Models
Martin Kuhn; Tony Trinh; Joscha Grüger; Ralph Bergmann
In: Roberto Posenato; Irene Vanderfeesten (Hrsg.). Advanced Information Systems Engineering Workshops. International Conference on Advanced Information Systems Engineering (CAiSE-2026), International Conference on Advanced Information Systems Engineering, located at CAiSE-2026, June 8-12, Verona, Italy, Pages 12-24, ISBN 978-3-032-28160-9, Springer Nature Switzerland, 2026.
Zusammenfassung
Data-driven process analytics, including machine learning and process mining, rely on event logs that capture how business processes are executed. In many organizations, these logs are sensitive, incomplete, or difficult to share. This has increased interest in synthetic event log generation, especially conditional generation that creates synthetic logs under user-defined constraints. Thus, we propose the Petri Net Conditioned Synthesizer (PNC-SYN) to generate synthetic logs that are conformant with a given normative Petri net. At each step, an autoregressive model predicts a probability distribution over the next possible events, and a Petri net playout layer restricts this distribution to the transitions enabled by the Petri net. Furthermore, the autoregressive model remains responsible for generating realistic event sequences and attributes such as timestamps, costs, and resources. We study (1) how to integrate Petri net playout with an autoregressive model to ensure conforming traces while enabling attribute generation and (2) how well Petri net conditioned synthetic logs match the original attribute distributions as well as the utility for process mining tasks compared to an unconstrained autoregressive synthesizer.
Projekte
- KI-AIM - KI-basierte Anonymisierung in der Medizin
- KI-AIM2 - KI-basierte Anonymisierung in der Medizin 2