Social News Website Moderation through Semi-supervised Troll User Filtering

Recently, Internet is changing to a more social space in which all users can provide their contributions and opinions to others via websites, social networks or blogs. Accordingly, content generation within social webs has also evolved. Users of social news sites make public links to news stories, so that every user can comment them or other users’ comments related to the stories. In these sites, classifying users depending on how they behave, can be useful for web profiling, user moderation, etc. In this paper, we propose a new method for filtering trolling users. To this end, we extract several features from the public users’ profiles and from their comments in order to predict whether a user is troll or not. These features are used to train several machine learning techniques. Since the number of users and their comments is very high and the labelling process is laborious, we use a semi-supervised approach known as collective learning to reduce the labelling efforts of supervised approaches. We validate our approach with data from ‘Menéame’, a popular Spanish social news site, showing that our method can achieve high accuracy rates whilst minimising the labelling task.