Representative sequence selection in unsupervised anomaly detection using spectrum kernel with theoretical parameter setting

摘要

Unsupervised anomaly detection is an important topic of data mining research, especially with respect to non-numerical sequence data. However, the majority of previous algorithms features empirical parameter selection. The contribution of this study is twofold: First, we show how the Akaike Information Criterion can be used to set the parameter of the spectrum kernel. Second, a distance-based algorithm for one-class unsupervised anomaly detection is presented. The algorithm uses the distance matrix of the data to select a sequence representative of the normal class by means of robust statistics. The proposed algorithm is applied to two kinds of sequence data, showing its suitability.