Business Process Management Anomaly Detection Through Semantic Embedding-Integrated Graph Neural Networks

Detecting anomalies in business operations is paramount to maintaining business efficiency and fulfilling operational goals. Anomalies in business processes can pose major threats to a company's operations; therefore, their detection is crucial in business process management (BPM) tools. While several current methodologies predominantly rely on statistical frequency for anomaly detection, it is critical to recognize that infrequency does not inherently mean undesirability. The complexities of business processes, along with the scarcity of labeled data, make successful anomaly detection even more difficult, necessitating novel methodologies and advanced algorithms. This study proposes a novel, fully unsupervised graph neural network architecture integrated with natural language processing capabilities, designed for scalability and applicability to real-world business process logs for anomaly detection. Our methodology employs graph autoencoders with edge-conditioned convolutions in conjunction with sentence transformers to enhance the analytical depth of business process state-to-state relationships by incorporating the semantic meaning of the processes. To validate the efficacy and versatility of our model, we conduct empirical experiments utilizing three distinct real-life logs, demonstrating its applicability in both trace-level and event-level anomaly detection scenarios.