Process Comparison Based on Selection-Projection Structures

Insight into differences between different implementations of a process provides valuable information for improvement. Process comparison approaches leverage event data on process executions to provide such insight. However, state-of-the-art procedural methods are often limited to local differences considering activities executed within a limited number of steps (e.g., directly following activities). Thereby, detecting differences which, for instance, relate early steps of a process execution to its outcome remains challenging. In contrast, rule-based declarative approaches can detect global differences with respect to distant activities; yet they are limited by the complexity of the rule templates employed . Moreover, they are prone to yield fragmented diagnostics. If a subprocess occurs more frequently in one process variant, these approaches typically report each activity contained. In this work, we therefore propose a process comparison approach that detects aggregated likelihood differences for global control-flow patterns. To this end, we decompose the difference detection task into subprocesses induced by co-occurring activities. Using Earth Mover’s Distance, we identify differences within individual subprocesses independent of predefined rule templates. We then aggregate and combine subprocesses which distinguish the process variants. By exploiting relations among subprocesses, we retrieve maximal differences affecting many activities. Reducing fragmentation caused by choice-induced frequency differences, we additionally complement these maximal differences. To compare the sensitivity of our difference detection method to existing approaches, we devise a quantitative evaluation framework. Moreover, we demonstrate the effectiveness of our method on a public, real-life event log. Ultimately, the evaluation shows that our method is accurate and capable of providing coherent, global diagnostics.