Identifying Rebel Twitter Users using Semi-Supervised Machine Learning

Abstract

Miscreant users use social media for attaining their purposes like recruitment, fundrais ing, and spreading propaganda. Among diverse types of miscreant social media users, majority of the researchers have focused on identifying users associated with terror ism and limited research is conducted for identifying rebel users. The focus of this thesis is to identify rebel users on social media. As many rebel users share strong sentiments, therefore it is important to understand these sentiments. We propose a novel sentiment classifier named Context-aware Sliding Window (CSW), which uses clues from the past sentiments to classify the sentiments of the current social media post. We also develop a new temporally labeled sentiment dataset. The result shows that the proposed CSW framework has improvements in temporal window size less than 24 hours and k less than 4. The proposed sentiment classification approach along with other proposed techniques is then used in identifying rebel users. For this purpose, we propose a Supervised Rebel Identification (SRI) framework which uses a novel directed user graph as a feature of the framework. We convert the user graph into graph embeddings which models the contexts of the users in lower dimensions. For evaluation, we develop a first multicultural and multiregional dataset representing rebel users affiliated with nine rebel movements across five countries. We compare the proposed SRI framework with baselines, including the state-of-the-art document embeddings. The proposed SRI model achieved near to 100% F1-score. This means the model can understand the context of a user in a better way. Finally we propose vi a semi-supervised rebel identification (SSRI) frameworkto identify new rebel users. The proposed SSRI model is compared against the proposed SRI model. The gap for improvement is minute because the SRI model has F1-score near 100%. However, the proposed SSRI model achieved more than 90% F1-score, which is acceptable because adding new users have an effect on accuracy. Still, the SSRI model is able to identify unknown or new rebellious users.