Multivariate Anomaly Detection in Object-Centric Event Data

Statistical anomaly detection is a powerful tool for identifying irregularities and improvement potential in process execution data. Existing research in this area has predominantly focused on applications in case-centric event data, considering a limited set of attributes. In this paper, we systematically study event-level anomaly detection in an object-centric context. We define and categorize the anomalies that can occur and derive detection strategies and features for object-related anomalies. We also introduce two novel object-centric datasets with labeled anomalies of varying complexity and empirically evaluate an existing detection approach on them. We find that methods developed for case-centric event logs can be effective in identifying complex and object-related anomalies, but suffer from robustness issues.