LSTM-Based Anomaly Detection of Process Instances: Benchmark and Tweaks

Anomaly detection can identify deviations in event logs and allows businesses to infer inconsistencies, bottlenecks, and optimization opportunities in their business processes. In recent years, various anomaly detection algorithms for business processes have been proposed based on either process discovery or machine learning algorithms. While there are apparent differences between machine learning and process discovery approaches, it is often unclear how they perform in comparison. Furthermore, deep learning research in other domains has shown that advancements did not solely come from improved model architecture but were often due to minor pre-processing and training procedure refinements. For this reason, this paper aims to set up a broad benchmark and establish a baseline for deep learning-based anomaly detection of process instances. To this end, we introduce a simple LSTM-based anomaly detector utilizing a collection of minor refinements and compare it with existing approaches. The results suggest that the proposed method can significantly outperform the existing approaches on a large number of event logs consistently.