Mining usage scenarios in business processes: Outlier-aware discovery and run-time prediction

A prominent goal of process mining is to build automatically a model explaining all the episodes recorded in the log of some transactional system. Whenever the process to be mined is complex and highly-flexible, however, equipping all the traces with just one model might lead to mixing different usage scenarios, thereby resulting in a spaghetti-like process description. This is, in fact, often circumvented by preliminarily applying clustering methods on the process log in order to identify all its hidden variants. In this paper, two relevant problems that arise in the context of applying such methods are addressed, which have received little attention so far: (i) making the clustering aware of outlier traces, and (ii) finding predictive models for clustering results. The first issue impacts on the effectiveness of clustering algorithms, which can indeed be led to confuse real process variants with exceptional behavior or malfunctions. The second issue instead concerns the opportunity of predicting the behavioral class of future process instances, by taking advantage of context-dependent "non-structural" data (e.g., activity executors, parameter values). The paper formalizes and analyzes these two issues and illustrates various mining algorithms to face them. All the algorithms have been implemented and integrated into a system prototype, which has been thoroughly validated over two real-life application scenarios.