Semi-Supervised Detection of Contaminated Business Process Instances Using Graph Autoencoders and Dynamic Edge Convolutions for BPM Anomaly Detection

Anomalies in business processes pose a significant threat to the operational performance of modern enterprises. Consequently, detecting these anomalies in business process management (BPM) systems is a critical task for organizations. However, the inherent complexity of business processes and the limited availability of labeled data compound this challenge. In this study, we introduce innovative models that leverage graph representations of business process data, incorporating an unsupervised neural network architecture known as Graph Autoencoders combined with Dynamic Edge Conditioned Convolutions to enhance learning on edge attributes. Additionally, we propose a novel metric for evaluating autoencoder performance. We developed two models: a fully unsupervised model trained on the entire dataset and a semi-supervised model that refines the initial model's outputs. Our empirical results demonstrate that the unsupervised model achieves a maximum F1-score of 0.82 and a success rate of 74.11% at the 80th percentile threshold for detecting anomalous sub-sequences in business processes. The semi-supervised model, which builds on the unsupervised model's findings, achieves a maximum F1-score of 0.89 and a success rate of 80.50%, indicating a substantial performance improvement. Both models offer promising capabilities for automating anomaly detection in BPM systems.