Building instance graphs for highly variable processes

Organizations increasingly rely on business process analysis to improve operations performance. Process Mining can be exploited to distill models from real process executions recorded in event logs, but existing techniques show some limitations when applied in complex domains, where human actors have high degree of freedom in the execution of activities thus generating highly variable processes instances. This paper contributes to the research on Process Mining in highly variable domains, focusing on the generation of process instance models (in the form of instance graphs) from simple event logs. The novelty of the approach is in the exploitation of filtering Process Discovery (PD) techniques coupled with repairing, which allows obtaining accurate models for any instance variant, even for rare ones. It is argued that this provides the analyst with a more complete and faithful knowledge of a highly variable process, where no process execution can be really targeted as "wrong" and hence overlooked. The approach can also find application in more structured domains, in order to obtain accurate models of exceptional behaviors. The quality of generated models will be assessed by suitable metrics and measured in empirical experiments enlightening the advantage of the approach.