A Benchmark of Process Drift Detection Tools: Experimental Protocol and Statistical Analysis.

Business processes are sequences of activities performed to achieve a specific goal, e.g., applying a clinical protocol to a patient. Process mining provides tools and techniques for analyzing and enhancing business processes. However, these processes are usually dynamic and can change because of new regulations, emergencies, or other reasons; and these changes are named concept or process drifts. Detecting drifts allows managers to improve the process analysis and act proactively upon these changes. We benchmarked three process drift detection tools and compared them based on an experimental protocol designed to evaluate the accuracy of the detected drifts. The selected tools detect sudden drifts in event logs. The experimental protocol generated a dataset containing the accuracy metric and the parameter configuration applied in each scenario. We applied statistical tests to verify significant differences in the accuracy between the tools when performed using distinct parameter configurations. The findings indicate that the parameter configuration affects the accuracy of the detected drifts and the dataset configuration. Another contribution of this paper is the designed experimental protocol, which can be applied to objectively evaluate the process drift detection tools and the dataset containing the results of the accuracy calculated over the performed experiments.