A Framework for Online Process Concept Drift Detection from Event Streams

Business processes are prone to variations due to the phenomenon known as concept drift, which refers to the situation where the business process is changing while being executed. Concept drift detection of business processes is one of the most critical problems in process mining. Most of the state-of-the-art approaches use statistical hypothesis testing for change point detection by sliding-window-based analysis and feature extraction, which suffer from bad performance and missed detection of certain types of changes. To attack such challenges, this paper proposes a framework as well as detailed models and techniques for online process concept drift detection without any feature extraction. A streaming scheme is presented to detect, locate, and classify the concept drifts simultaneously without requiring any storage of the event traces that have been analyzed. Four types of process concept drifts which are sudden drift, gradual drift, recurring drift, and incremental drift can be precisely discovered, and noise as well as flexibility in business processes are fully considered. The efficacy of the approach is also validated by simulation experiments on large-scale business process event streams.