IMPORTANCE OF SCALING IN UNSUPERVISED DISTANCE-BASED ANOMALY DETECTION
Pekka S. Kumpulainen, Mikko Kylväjä, Kimmo Hätönen · 2009
Abstract − One of the key applications in mobile network monitoring is to detect anomalous phenomena in the network. Distance-based methods are commonly used in unsupervised anomaly detection. The results are dependent on the distance metrics used and the scaling of the variables. In many cases very simple methods can provide sufficient performance if the variables have been scaled properly. In this paper we discuss the importance of scaling in distancebased methods and the possibility to incorporate a priori knowledge of the relative importance of the variables by scaling. We present an example of a priori scaling on performance data measured from the radio interface in a mobile telecommunication network. The results are compared to those obtained by using traditional normalization.