Business processes constantly evolve as organizations adapt to new regulations, technologies, and market conditions. Consequently, models discovered through process mining quickly lose validity unless process concept drifts are detected and managed. The newest and best drift detection approaches rely on supervised learning or complex statistical models, demanding labeled data and high computational effort. This paper introduces Version Clustering (VC), a novel top-down unsupervised approach for process concept drift detection that is both robust and interpretable. VC's novelty lies in the inversion of the drift detection logic in comparison to existing approaches. Rather than pinpointing drift points outright, VC first clusters windows into coherent process versions via hierarchical density-based clustering, and then marks drift points where execution transitions between versions. This inversion yields three advantages: (i) improved robustness to noise and local behavioral variability, (ii) more accurate predictions, and (iii) a version-level representation that naturally supports visual and analytical interpretation. Evaluated within a unified experimental framework and tested on established benchmarks, VC achieves high robustness to noise, strong accuracy, practical scalability, coverage of most drift types, and clear interpretability. Under the considered evaluation metrics, it matched or exceeded baseline approaches, offering an efficient and interpretable solution for tracking process evolution in noisy, dynamic environments.

Top