Multi-Aspect Anomaly Detection with Graph Neural Networks and Kolmogorov-Arnold Networks in Business Process Management

Anomalous occurrences within business processes can have a significant negative impact onab usiness lifecycle. When the level of competitiveness found in modern-day markets is high, an anomaly ranging anywhere from a basic inefficiency to fraudulent behavior can deteriorate a business's profitability, putting it at a disadvantage in front of its competitors. The digitalization of the modern world and the widespread availability and usage of business process management (BPM) solutions enable the detection of these anomalies through the use of business process event logs. In this study, we leveraged the graph representation of business process traces by using a Graph Autoencoder (GAE) built by using Edge-Conditioned Convolutions (ECC) and Kolmogorov-Arnold Networks (KAN) for improvement over models built with multi-layer perceptrons (MLP) to act as the decoder. In addition to using more advanced neural network architectures, we built one dimension increasing and one standard autoencoder (AE) to see if dimension increasing AEs would perform better regarding an anomaly detection task. The empirical results show that leveraging a KAN with a standard GAE yields an f'l-score of 0.50, an increase of 0.08 in f'l-score when put side by side with the GAE using the MLP decoder, which yields an f'l-score of 0.42 on edge labels. On the trace level, the standard model with the KAN decoder yielded a 0.03 increase in f'l-scorc from 0.67 to 0.70 in comparison to the MLP model. The results show us that leveraging KANs and GNNs together can make for enhanced anomaly detection on business process traces.