Enhancing explainability in process variant analysis: a framework for detecting and interpreting control-flow changes

Abstract Processes often exhibit significant variability, posing challenges for process discovery and insight extraction. While most studies focus on detecting variability over time (e.g., concept drift), control-flow variability can also manifest across other dimensions, such as case durations or performance metrics. Identifying and understanding these changes is vital for uncovering inefficiencies and undesired behaviors. This paper introduces a novel framework that combines control-flow change detection across performance dimensions with explainability, providing insights into where and how control flow evolves. The framework uses a sliding window approach with the earth mover’s distance to detect behavioral shifts. To enhance interpretability, event logs are encoded into a feature space defined by declarative constraints, capturing intuitive control-flow properties. Clustering these features reveals distinct behavioral patterns and their evolution along performance dimensions, linking detected changes to specific process dynamics. We validate the framework using three real-life event logs, including one from the UWV employee insurance agency in the Netherlands, demonstrating its ability to uncover meaningful changes, explain process variability, and support data-driven decision-making. The framework is implemented as an open-source tool for broader applicability.