The Metadata Troll Detector

Reddit.com offers a discussion platform for various topics. Given its huge size and its comment based discussion structure it attracts internet trolls. We write a bot that crawls Reddit for new comments and that tries to automatically detect trolling in them. In contrast to conventional solutions we do not only use text analysis of the comments but also take metadata into account. Metadata describes properties such as the degree of the participation in a discussion. For classification, we compile several characteristics of a comment based on data and metadata. These characteristics are used to train the machine learning algorithms of support vector machines to classify each comment. We show that it is generally very difficult to automatically detect trolls and see that metadata based approaches are still inferior to text based ones. However, the combination of both shows promising results.