Detecting Malicious Users on Twitter Using Topic Modeling

Abstract

Social networking sites like Twitter, Weibo and Facebook etc. have played a significant role in daily lives of people as they promote innovative ways to connect efficiently and exchange their knowledge. The benefits of these social network services cause them to expand their community rapidly. Most current social network sites face additional issue of coping with unauthorized users and their high levels of violence activities, which distribute fake news, worms and viruses, etc. to the genuine users. Spam distribution degrades user experience and also has a negative effect on server-side functions such as knowledge discovery, user activity analysis and service selection. In this paper, whitelist and blacklist are built which can help to distinguish malicious users and legitimate users. With the aid of blacklist and whitelist, we introduced two new features: malicious probability and legitimate probability. Evaluation has been carried out on the CRESCI-2015 dataset. Three machine learning classifiers like AdaBoost, Bagging and Random Forest. Random Forest obtained the highest 99.7% detection score.