Multi-Graph Anomaly Detection in Business Processes

This technical report outlines the challenges associated with business process anomaly detection and proposes a scalable and efficient solution based on enhanced graph-based autoencoder models. The study focuses on developing a production-ready anomaly detection system that is adaptable across various business domains. It details improvements made to artificial anomaly injection methods to better represent real-world scenarios, addressing the scarcity of annotated datasets. A comparative analysis of multiple model architectures is presented, highlighting the use of Graph Attention v2 and Transformers to achieve significant performance outcomes with reduced computational costs. The report emphasizes the importance of reliable detection capabilities in different operational contexts, outlining criteria for effective anomaly detection systems, including high precision and recall, adaptability to diverse domains, and efficient performance. The results demonstrate notable improvements in scalability and detection accuracy, indicating the proposed solution’s robustness in identifying anomalies within business process workflows.