Machine learning-based detection of concept drift in business processes

Abstract Concept drift in process mining occurs when a single event log includes data from multiple versions of a process, making the detection of such drifts essential for ensuring reliable process mining results. Although many techniques have been proposed, they exhibit limitations in accuracy and scope. Specifically, their accuracy diminishes when facing noise, varying drift types, or different levels of change severity. Additionally, these techniques primarily focus on detecting sudden and gradual drifts, overlooking the automated detection of incremental and recurring drifts. To address these limitations, we present , a novel approach for automated concept drift detection that can identify sudden, gradual, incremental, and recurring drifts. Our approach follows an entirely different paradigm. Specifically, it employs a supervised machine learning model fine-tuned on a large collection of event logs with known concept drifts, enabling the model to learn how drifts manifest in event logs. The possibility to train such a model has recently emerged through a tool that generates event logs with known concept drifts. However, applying supervised machine learning remains challenging due to the complexities of encoding. To address this, we propose converting an event log into an image-based representation that captures process evolution over time, enabling the use of a state-of-the-art computer vision model to detect drifts. Our experiments show that our approach, compared to existing solutions, improves the accuracy and robustness to noise of drift detection while extending coverage to a broader range of drift types, highlighting the potential of this new paradigm.