How Object-Centric is Object-Centric Predictive Process Monitoring? Introducing Objects into Object-Centric Predictive Process Monitoring

Abstract

The field of process mining is challenged by the complexity of true processes when extracting accurate process behavior and statistics from information systems. Traditional process mining algorithms assume a single case notion, whereas actual processes contain many possible ones, represented by objects. Object-centric process mining has been introduced as a case-agnostic solution, which mitigates the problem of misleading process behavior and statistics via the object-centric event log (OCEL). It allows multiple case notions which are called object types. Objects have many-to-many relationships with events. However, when performing predictive process monitoring on OCELs, issues arise when including object information as features due to these many-to-many relationships. This has not been addressed by existing literature. We propose a heterogeneous object event graph encoding (HOEG), that incorporates events and objects into a graph with different node types. We evaluate our novel encoding against an extant graph-based encoding and several baselines on the task of remaining time prediction. On our HOEG we employ a heterogeneous graph neural network (GNN) architecture that is converted from a homogeneous one. The HOEG-based GNN learns an optimal way to include object information when forming predictions. The experiments are executed on three OCELs, one of which is extracted from an operational process at a large Dutch financial institution. Our results indicate that HOEG outperforms its competition for well-structured OCELs. Furthermore, we argue that HOEG mainly excels when OCELs host informative object attributes and abundant object interactions. Considering this, we propose HOEG as a promising general technique to leverage the multi-dimensional data structure given in OCELs for tasks like predicting process remaining time.