Change point detection and dealing with gradual and multi-order dynamics in process mining

In recent years process mining techniques have matured. Provided that the process is stable and enough example traces have been recorded in the event log, it is possible to discover a high-quality process model that can be used for performance analysis, compliance checking, and prediction. Unfortunately, most processes are not in steady-state and process discovery techniques have problems uncovering "second-order dynamics" (i.e., the process itself changes while being analyzed). This paper describes an approach to discover a variety of concept drifts in processes. Unlike earlier approaches, we can discover gradual drifts and multi-order dynamics (e.g., recurring seasonal effects mixed with the effects of an economic crisis). We use a novel adaptive windowing approach to robustly localize changes (gradual or sudden). Our extensive evaluation (based on objective criteria) shows that the new approach is able to efficiently uncover a broad range of drifts in processes.