Semi-automated Time-Granularity Detection for Data-Driven Simulation Using Process Mining and System Dynamics

Most information systems supporting operational processes also record event logs. These can be used to diagnose performance and compliance problems. The majority of process mining techniques extract models that are descriptive and describe what happened in the past. Few process mining techniques discover models that allow us to ""look into the future"" and perform predictive analyses. Recently, novel approaches have been developed for scenario-based prediction, i.e., predicting the effects of process changes on process performance, e.g., investing in an additional resource. To work accurately, the techniques need an appropriate time step-size, the selection of which, thus far, has been an ad-hoc and manual endeavor. Therefore, in this paper, building upon time-series analysis and forecasting techniques, we propose a novel semi-automated time-granularity detection framework. Our framework detects the best possible time-granularity to be used, whilst taking user preferences into account. Our evaluation, using both real and synthetic data, confirms the feasibility of our approach and highlights the importance of using accurate granularity in time step selection.