Semi-supervised Detection of Business Process Anomalies

This technical report presents a novel approach to detect anomalies in business process management (BPM) systems using semi-supervised learning techniques. The study introduces two models: a fully unsupervised model and a semi-supervised model, both leveraging graph autoencoders and dynamic edge convolutions. The unsupervised model achieves a maximum F1-score of 0.82, while the semi-supervised model improves this to 0.89, demonstrating enhanced performance in detecting anomalous sub-sequences in business processes. The report outlines the challenges of anomaly detection, including the complexity of business processes and the scarcity of labeled data. It emphasizes the importance of automating anomaly detection to maintain operational efficiency and prevent fraudulent activities. The methodology section details the model construction and dataset used, which consists of over 44,000 rows of business process logs. The findings indicate that the proposed models can effectively capture complex relationships in business process data, contributing to the field of BPM anomaly detection.