Extracting Process Features from Event Logs to Learn Coarse-Grained Simulation Models

Most process mining techniques are backward-looking, i.e., event data are used to diagnose performance and compliance problems. The combination of process mining and simulation allows for forward-looking approaches to answer "What if?" questions. However, it is difficult to create fine-grained simulation models that describe the process at the level of individual events and cases in such a way that reality is captured well. Therefore, we propose to use coarse-grained simulation models (e.g., System Dynamics) that simulate processes at a higher abstraction level. Coarse-grained simulation provides two advantages: (1) it is easier to discover models that mimic reality, and (2) it is possible to explore alternative scenarios more easily (e.g., brainstorming on the effectiveness of process interventions). However, this is only possible by bridging the gap between low-level event data and the coarse-grained process data needed to create higher-level simulation models where one simulation step may correspond to a day or week. This paper provides a general approach and corresponding tool support to bridge this gap. We show that we can indeed learn System Dynamics models from standard event data.