Accuracy of statistical machine learning methods in identifying client behavior patterns at network edge

Abstract

This paper is focused on evaluating the applicability of statistical machine learning methods to identifying flows and user behavior patterns at the source (client) network edge. The research was conducted in a mid-size (covering ca 150 geographically scattered locations) network developed for the Malopolska Educational Cloud (MEC) project. Due to the lack of validation sets we focused on unsupervised learning methods. Modules implementing the methods were fed with the headers of the user-generated packets; payloads were not analyzed due to privacy concerns. The presented research proved that in client edge networks even the simple classification methods yield satisfactory results in flows classification.