Publication details
This post summarizes the peer-reviewed work cited below.
- Paper
- Business Process Management Anomaly Detection Through Semantic Embedding-Integrated Graph Neural Networks
- Authors
- Teoman Berkay Ayaz; Ege Gülce; Stanley Hsu; Alper Özcan; Akhan Akbulut
- Presented at
- 2024 9th International Conference on Computer Science and Engineering (UBMK), 26–28 October 2024, Antalya, Türkiye
- Pages
- 551–556
- Publication date
- October 26, 2024
- DOI
- 10.1109/ubmk63289.2024.10773613 — Publisher page
Deviations in business processes can directly threaten an organization's operational goals, which is why anomaly detection is a critical capability in Business Process Management (BPM) tools. Most existing approaches look at how rarely a behavior occurs; yet not everything that is rare is faulty. The complexity of processes and the scarcity of labeled data make the task harder still.
This joint study by researchers from the Next4biz R&D Center and Akdeniz University approaches the problem with a fully unsupervised graph neural network architecture. The model represents the transitions between process steps as a graph; it combines a graph autoencoder that uses edge-conditioned convolutions with sentence transformer embeddings that capture the meaning of the steps. In this way, the analysis covers not only the information of "which step was followed by which," but also what those steps mean. The architecture was designed to be applied to real-world process logs in a scalable way.
The method was tested on three different real-life process logs, and it was shown that it can be used for anomaly detection both at the case (trace) level and at the event level.
The paper was presented at the 9th International Conference on Computer Science and Engineering (UBMK 2024), held in Antalya on 26–28 October 2024, and it appeared on pages 551–556 of the conference proceedings published by IEEE. The study focuses on the same research topic as the AnomalyNet project that the R&D team carries out for Next4biz BPM: detecting process anomalies with graph neural networks.
