Anomaly Detection Method for Business Processes Based on Graph Convolutional Neural Network

Anomaly detection in business processes enables real-time monitoring, which helps enterprises intervene before issues arise to ensure process security and continuity. This paper proposes an anomaly detection method based on Graph Convolutional Networks (GCN). We extract prefix traces from offline event logs to build a dataset with anomaly labels. Local and global features of these traces are embedded using a combination of Word2Vec and GCN, with the final embeddings generated through their fusion. Anomalies are then detected using neural network classifiers. Experiments on four real event log datasets show that our model improves detection accuracy compared to BERT and Transformer algorithms, achieving a peak accuracy of 93 % and an average accuracy of 92%.