Anomaly detection from telecommunication data using three-way analysis

Abstract

So far, several supervised and unsupervised solutions have been provided for detecting failures in telecommunication networks. Among them, unsupervised approaches attracted more attention since no labeled data is required. Principal component analysis (PCA) is a wellknown unsupervised technique to solve this type of problem when data is organized in matrix form. However, PCA may fail to capture all the significant interactions established among di erent dimensions when applied to higher-order data. When dealing with three, instead of two dimensions, three-way factorization methods are more suitable since they are able to explicitly take into account the interactions among the three dimensions, without collapsing the raw data. Since an important property of telecommunication data is its temporal-sequential nature, the temporal dimension should be considered along with the other two dimensions in order to gain insights regarding its evolution. The aim of this paper is to demonstrate t