Graph Autoencoders for Business Process Anomaly Detection

We propose an approach to identify anomalies in business processes by building an anomaly detector using graph encodings of process event log data coupled with graph autoencoders. We evaluate the proposed approach with randomly mutated real event logs as well as synthetic data. The evaluation shows significant performance improvements (in terms of F1 score) over previous approaches, in particular with respect to other types of autoencoders that use flat encodings of the same data. The performance improvements are also stable under training and evaluation noise. Our approach is generic in that it requires no prior knowledge of the business process.