Unsupervised Anomaly Detection in Noisy Business Process Event Logs Using Denoising Autoencoders

Abstract

Business processes are prone to subtle changes over time, as unwanted behavior manifests in the execution over time. This problem is related to anomaly detection, as these subtle changes start of as anomalies at first, and thus it is important to detect them early. However, the necessary process documentation is often outdated, and thus not usable. Moreover, the only way of analyzing a process in execution is the use of event logs coming from process-aware information systems, but these event logs already contain anomalous behavior and other sorts of noise. Classic process anomaly detection algorithms require a dataset that is free of anomalies; thus, they are unable to process the noisy event logs. Within this paper we propose a system, relying on neural network technology, that is able to deal with the noise in the event log and learn a representation of the underlying model, and thus detect anomalous behavior based on this representation. We evaluate our approach on five different event logs, coming from process models with different complexities, and demonstrate that our approach yields remarkable results of 97.2 % F1-score in detecting anomalous traces in the event log, and 95.6 % accuracy in detecting the respective anomalous activities within the traces.