Business Process Anomaly Detection with Sequence-Capturing Hybrid Autoencoders

In deep learning, every network architecture has its own strengths and shortcomings. One way to bring those strengths together is to combine different neural networks into a hybrid model. Business process anomaly detection is one of the domains where specialized neural networks can yield a considerable performance increase. The paper by Teoman Berkay Ayaz, Alper Özcan and Akhan Akbulut explores hybrid architectures in this domain.

The study runs sequence-capturing autoencoders in tandem with convolutional neural networks (CNN). The autoencoders extract feature-rich latent representations from business process sequences; the CNN layers enrich those representations with additional feature extraction.

The results show that combining different architectures yields performance increases of up to 8% in event-level anomaly detection (from 0.552 to 0.633 in F1-score) and up to 3% in trace-level anomaly detection (from 0.734 to 0.766). The hybrid model is benchmarked against a GRU-only autoencoder on five datasets released as part of the BPI Challenge 2020, and it does not come out ahead on all of them: the gain shrinks as attribute dimensionality grows, and on the BPIC20-PR dataset the hybrid model falls behind at both the event and trace levels (F1-score from 0.390 to 0.365 and from 0.617 to 0.555). The paper also reports additional empirical findings on hybrid deep model construction.

The paper was presented at ISAS 2025 (9th International Symposium on Innovative Approaches in Smart Technologies), held in Gaziantep on June 27-28, 2025, and appears in the proceedings published by IEEE (pp. 1-5). For our journal article in the same field, see our post on SPECTRE.