Multi-perspective Identification of Event Groups for Event Abstraction

Abstract

AbstractIn process mining settings, events are often recorded on a low level and cannot be used for meaningful analysis directly. Moreover, the resulting variability in the recorded event sequences leads to complex process models that provide limited insights. To overcome these issues, event abstraction techniques pre-process the event sequences by grouping the recorded low-level events into higher-level activities. However, existing abstraction techniques require elaborate input about high-level activities upfront to achieve acceptable abstraction results. This input is often not available or needs to be constructed, which requires considerable manual effort and domain knowledge. We overcome this by proposing an approach that suggests groups of low-level events for event abstraction. It does not require the user to provide elaborate input upfront, but still allows them to inspect and select groups of events that are related based on their common multi-perspective contexts. To achieve this, our approach learns representations of events that capture their context and automatically identifies and suggests interesting groups of related events. The user can inspect group descriptions and select meaningful groups to abstract the low-level event log.