Looking for Change: A Computer Vision Approach for Concept Drift Detection in Process Mining

Concept drift in process mining refers to a situation where a process undergoes changes over time, leading to a single event log containing data from multiple process versions. To avoid mixing these versions up during analysis, various techniques have been proposed to detect concept drifts. Yet, the performance of these techniques, especially in situations when event logs involve noise or gradual drifts, is shown to be far from optimal. A possible cause for this is that existing techniques are developed according to algorithmic design decisions, operating on assumptions about how drifts manifest themselves in event logs, which may not always reflect reality. In light of this, we propose a completely different approach, using a deep learning model that we trained to learn to recognize drifts. Our computer vision approach for concept drift detection (CV4CDD) uses an image-based representation that visualizes differences in process behavior over time, which enables us to subsequently apply a state-of-the-art object detection model to detect concept drifts. Our experiments reveal that our approach is considerably more accurate and robust than existing techniques, highlighting the promising nature of this new paradigm.