Business processes are subject to changes during their execution over time due to new legislation, seasonal effects, etc. Detection of process changes is alternatively called business process drift detection. Currently, existing methods unfavorably subject the accuracy of drift detection to the effects of window size. Furthermore, most methods have to struggle with the problem of selecting appropriate features specifying the relations between traces or events. This paper draws on the notion of trace embedding to propose a new framework for automatic detection of suddenly occurring process drifts. The main contributions of the proposed approach are: i) It is independent of windows; ii) Trace embedding that is used for drift detection makes it possible to automatically extract all features from relations among traces; iii) As attested by synthetic event logs, this approach is superior to current methods in terms of accuracy and drift detection delay.

Keywords