A Comparative Evaluation of SOM-based Anomaly Detection Methods for Multivariate Data

Anomaly detection for multivariate data is of vital importance in academic research and industry. In real scenes, there is usually a lack of labels of anomalies. Self-Organizing Map (SOM) can map data to the output layer and maintain the original topology, which has been used as a semi-supervised learning method to solve the above problem. In this paper, we first explain the mechanism of classic SOM for anomaly detection, then compare it with two variants of SOM named kernel SOM and K-BMUs SOM. Kernel SOM replaces Euclidean distance with kernel functions, while K-BMUs SOM changes the number of matching neurons. The three types of SOM are applied to multivariate datasets in three different domains. We find that the performance of the three SOM-based methods is related to the characteristics of data.