Examining the performance of classification algorithms for imbalanced data sets in web author identification

Individuals, criminals or even terrorist organizations can use web-communication for criminal purposes; to avoid the prosecution they try to hide their identity. To increase level of safety in Web we have to improve the author (or web-user) identification and authentication procedures. In field of web author identification the situation of imbalanced data sets appears rather frequent, when number of one author's texts significantly exceeds the number of other's. This is common situation for the modern web: social networks, blogs, emails etc. Author identification task is some sort of classification task. To develop methods, technics and tools for web author identification we have to examine the performance of classification algorithms for imbalanced data sets. In this work several modern classification algorithms were tested on data sets with various levels of class imbalance and different number of available webpost The best accuracy in all experiments was achieved with Random Forest algorithm.

I. Introduction

In modern Web millions of new electronic texts appears every day. Web-users identification and authentication procedures are not enough effective; users could produce numerous of identities, present a fake identity or even hide it [1]. Author identification can be used in various tasks in field of intelligence, computer forensics and in cyber (or on-line) security to determine the identity of the author based on analyses of his texts [2]–[4].