Concept drift detection using SGT and trace clustering in process mining

Abstract

Business processes change and adapt to the environment over time. So, detecting the exact point of change in the business process, also called concept drift, is essential for improving the process performance. However, the existing approaches have problems with the detection accuracy, reflection of feature information, and robustness on an anomaly. This paper proposes a new method for detecting concept drift, with high robustness to an anomaly and high accuracy using new features. Especially, our method better recognizes some specific patterns such as loops or concurrencies, reflects the overall trace flow, and takes into account the anomaly in the business process. These features prevent false positives or misses in detection of drift, resulting in higher accuracy. Experiments on synthetic and real-life logs were conducted, with comparison to the previous methods, showing a comparable performance on the synthetic log and the best performance on finding the validated drift points on the real-life log.