Detecting and Coloring Anomalies in Real Cellular Network Using Principle Component Analysis

摘要

Anomaly detection in a communication network is a powerful tool for predicting faults, detecting network sabotage attempts and learning user profiles for marketing purposes and quality of services improvements. In this article, we convert the unsupervised data mining learning problem into a supervised classification problem. We will propose three methods for creating an associative anomaly within a given commercial traffic data database and demonstrate how, using the Principle Component Analysis (PCA) algorithm, we can detect the network anomaly behavior and classify between a regular data stream and a data stream that deviates from a routine, at the IP network layer level. Although the PCA method was used in the past for the task of anomaly detection, there are very few examples where such tasks were performed on real traffic data that was collected and shared by a commercial company. The article presents three interesting innovations: The first one is the use of an up-to-date database produced by the users of an international communications company. The dataset for the data mining algorithm retrieved from a data center which monitors and collects low-level network transportation log streams from all over the world. The second innovation is the ability to enable the labeling of several types of anomalies, from untagged datasets, by organizing and prearranging the database. The third innovation is the abilities, not only to detect the anomaly but also, to coloring the anomaly type. I.e., identification, classification and labeling some forms of the abnormality.