ARIEN-supported research advances interpretable anomaly detection using Self-Organizing Maps

New research supported by the ARIEN project explores how unsupervised machine learning can improve anomaly detection while maintaining the interpretability needed in sensitive analytical contexts.

The paper, Unsupervised Anomaly Scoring Method based on Interpretation of Self-Organizing Maps, was authored by Niko Uremović and Niko Lukač from the University of Maribor. The accepted paper was presented at the Advanced Machine Learning and Data Science Conference – AMLDS 2026, held in Osaka, Japan, from 21 to 23 July 2026. The work acknowledges financial support from the Horizon Europe ARIEN project.

Moving beyond conventional anomaly scoring

Anomaly detection aims to identify observations that differ significantly from expected patterns. This can be particularly valuable in datasets where anomalous cases are rare and reliable labelled training data are difficult to obtain.

Self-Organizing Maps, or SOMs, are well suited to this type of unsupervised analysis because they transform complex data into lower-dimensional representations while retaining structural information. Their interpretability is especially valuable where analysts need not only to identify potentially unusual cases, but also to understand the patterns behind the results.

Many existing SOM-based anomaly detection methods rely primarily on quantization error. The new method proposed by the researchers instead combines this information with structural characteristics of the trained map.

The approach identifies dense neuronal kernels and broader communities within the SOM and considers factors such as the distance of neurons from these structures, neighbourhood divergence and activation frequency. These components are then combined with the individual sample’s quantization error to generate an anomaly score.

Testing the approach

The researchers evaluated the proposed method using four credit-card fraud datasets, including both real-world and synthetic transaction data. Performance was assessed using established anomaly-detection metrics and compared with conventional SOM scoring as well as other unsupervised methods.

The results show that incorporating structural information from the SOM consistently improved performance compared with quantization-error-only scoring. According to the study, the approach retains the interpretability advantages of Self-Organizing Maps while achieving performance comparable to, and in several cases exceeding, that of autoencoders and Isolation Forests.

Interpretability alongside performance

The work addresses a broader challenge in the application of Artificial Intelligence to complex analytical tasks: high-performing models are most useful in sensitive contexts when their outputs can also be inspected and understood by human operators.

By combining unsupervised anomaly detection with an interpretable representation of the underlying data structure, the research contributes to ARIEN’s wider exploration of AI-supported analytical approaches and human-centred use of advanced technologies.

The research received financial support from ARIEN – Artificial Intelligence in fighting illicit drugs production and traffickiNg, funded under the Horizon Europe programme, Grant Agreement No. 101121329.

📄 Read the paper on ARIEN Zenodo Community