Does This Make Sense? Machine Learning-Based Detection of Semantic Anomalies in Business Processes

Abstract

The detection of undesired behavior is a key task in process mining, supported by techniques for conformance checking and anomaly detection. A downside of conformance checking, though, is that it requires a process model as a basis, limiting its applicability, whereas existing anomaly detection techniques look for statistically infrequent behavior, even though infrequency does not necessarily imply undesirability. The recently introduced concept of semantic anomaly detection overcomes these issues by detecting behavior that stands out from a semantic point of view, such as a claim being paid after it has been rejected. In this manner, it detects behavior that is undesirable, while its grounding in natural language analysis allows it to consider behavioral regularities extracted from other processes, alleviating the need to have a process model available. However, the state-of-the-art approach for semantic anomaly detection, a rigid, rule-based approach, is limited in its scope and accuracy. Therefore, we propose a machine learning-based alternative, which uses a classifier trained to recognize whether observed process behavior is normal or anomalous. Our experiments show that this learning-based approach greatly outperforms the state of the art. Users can directly apply our approach to detect semantic anomalies in their own event data by using one of our pre-trained classifiers, even if their data contains so far unseen process behavior.