Revisiting the Transition Matrix-Based Concept Drift Approach: Improving the Detection Task Reliability Through Additional Experimentation

Coping with concept drifts in process mining remains highly pertinent, given the inherently dynamic essence of real-world business processes. Daily operations see processes undergoing continuous changes due to changing market demands, technological advances, or organizational restructuring. Conceptual drifts may emerge unexpectedly at any time, impacting process behavior significantly. Neglecting to manage these drifts can lead to obsolete process models, inaccurate performance analyses, and misguided decision-making. Researchers have been struggling to address several issues related to concept drifts in process mining; nevertheless, this question persists open. As a contribution, we previously proposed an approach based on transformed transition matrices as an efficient, simple, and extensible data structure, first applied to the task of detecting concept drifts. For this purpose, three strategies were adjusted to function with transformed transition matrices. We evaluated the approach’s effectiveness by initially using one set of event logs, contrasting its performance against cutting-edge benchmark methods in experimental settings. In this paper, we revisit the proposed approach, by expanding and reinforcing its evaluation through an additional set of event logs. The findings underscore the transformed transition matrix’s capability to encapsulate manifold features drawn from event logs, particularly those tied to process drifts via windowed comparisons. We show how these extracted features contribute to pinpointing sudden process drifts in control flow, specifically in an offline scenario. The experimental outcomes hold promise, signifying the potential of the three adapted detection strategies.