Detecting Political Bias Trolls in Twitter Data

Ever since Russian trolls have been brought to light, their interference in the 2016 US Presidential elections has been monitored and studied. These Russian trolls employ fake accounts registered on several major social media sites to influence public opinion in other countries. Our work involves discovering patterns in these tweets and classifying them by training different machine learning models such as Support Vector Machines, Word2vec, Google BERT, and neural network models, and then applying them to several large Twitter datasets to compare the effectiveness of the different models. Two classification tasks are utilized for this purpose. The first one is used to classify any given tweet as either troll or non-troll tweet. The second model classifies specific tweets as coming from left trolls or right trolls, based on apparent extreme political orientations. On the given data sets, Google BERT provides the best results, with an accuracy of 89.4% for the left/right troll detector and 99% for the troll/non-troll detector. Temporal, geographic, and sentiment analyses were also performed and results were visualized.