Identifying Execution States and Boundary Conditions in Object-Centric Event Logs: A Framework and Case Study

Abstract

Understanding how object-centric processes evolve requires discovering execution states directly from event data and explaining the conditions under which transitions occur. This paper introduces an unsupervised framework that infers states and their boundary conditions from sequences of events related to a business object. Each sequence is encoded as numerical vectors and projected into a low-dimensional space. On top of the projected points we train a self-organizing map (SOM) and associate each event, via its surrounding context, to a best-matching neuron, which we interpret as an execution state. Transitions arise from consecutive events mapping to different neurons, and boundary conditions are characterized by the event patterns and contextual signals that precede such changes. The approach uncovers non-obvious states that are not necessarily tied to any single data attribute, and it supports both within-state comparisons (which events tend to co-occur in a state) and between-state analyses (which event patterns drive movement across states). We provide a prototypal implementation and a case study in an inventory management context.

Authors 3

1. Alessandro Berti Aachen Process and Data Science (PADS)

   RWTH Aachen University

   Affiliation as printed

   Process and Data Science (PADS), RWTH Aachen University

2. Dina Kretzschmann Aachen Process and Data Science (PADS)

   RWTH Aachen University

   Affiliation as printed

   Process and Data Science (PADS), RWTH Aachen University

3. Wil M. P. van der Aalst Aachen Process and Data Science (PADS)

   RWTH Aachen University

   Affiliation as printed

   Celonis

   Process and Data Science (PADS), RWTH Aachen University