Multi-Aspect Anomaly Detection in BPM with Graph Neural and Kolmogorov-Arnold Networks

Deviations in business processes, ranging from simple inefficiency all the way to fraud, can damage an organization's profitability and competitiveness. As Business Process Management (BPM) solutions have become widespread, the process event logs they accumulate form fertile ground for catching these deviations automatically. This paper by the Next4biz R&D team targets anomaly detection over exactly these logs.

In the study, process traces are represented as graphs, and a graph autoencoder (GAE) built with Edge-Conditioned Convolution is used. The main novelty is the use of Kolmogorov-Arnold Networks (KAN) in the decoder layer instead of the customary multilayer perceptron (MLP). The team also compares two autoencoder variants, a dimension-increasing one and a standard one. The publisher's abstract does not state the source of the dataset used.

According to the results, the standard GAE with a KAN decoder raises the F1 score at the edge (step transition) level from 0.42 in the MLP-based model to 0.50, and at the trace level from 0.67 to 0.70. The authors conclude that bringing GNNs and KANs together strengthens process anomaly detection.

The paper was presented at UBMK 2024 (the 9th International Conference on Computer Science and Engineering), held in Antalya on October 26–28, 2024, and appeared in the proceedings published by IEEE (pp. 557–562). The authors are Teoman Berkay Ayaz, Ege Gülce, Stanley Hsu and Akhan Akbulut (Next4biz R&D Center) together with Alper Özcan (Akdeniz University).

Prof. Dr. Akhan Akbulut

Professor Doctor Akhan Akbulut worked in the Computer Engineering departments of Istanbul Kültür University and NC State University. He serves institutions such as TÜBİTAK, Ministry of Industry and Technology, TÜSEB, and KOSGEB. He researches Distributed Systems and Artificial Intelligence and has over 100 international journal articles and conference proceedings from his work.