A Framework for User Characterization based on Tweets Using Machine Learning Algorithms

Abstract

Twitter having more than three billion users is one of the most commercial and popular social networking sites. Twitter permits its users to post short messages and update their status. Tweets can be seen instantly by the followers of the users and other people with no twitter accounts. So by far most of the substance posted on the twitter is publicly accessible. Enormous number of political actors used twitter, who are interested in seeking extreme motives like radicalization, mobilization and recruiting activities. Twitters is used by large number of extremist organizations for press releases, public declaration and provide confirmation or motivation of their attacks. There have been several works looking at identifying extremist content based on twitter data but user identification using tweets has not been focused enough because of publication barrier and unavailability of data. In this research, a model is proposed which characterize a user into extremist and non-extremist categories. In this approach, pre-processing is done using natural language processing techniques and feature selection is performed using bag of words model. TF-IDF and word length is applied to obtain vector or feature to measure the significance of obtained vector in the whole document. We performed a methodology using classification through NB (Multinomial naïve Bayes) naïve Bayes on crises related tweets and Kaggle dataset related to tweets published by several Islamic State of Iraq and Sham to validate our proposed model. In this paper, a novel method is discussed for user characterization based on tweets posted by them. Evaluation results show that our suggested method gives best retrieval accuracies for word length feature extraction approach.