DeTroll—Leveraging Graph Neural Networks with Attention Mechanism to Detect State-Sponsored Trolls

Trolling on social media refers to the act of deliberately posting messages, comments, or content on social media platforms to cause a disruption or provoke a reaction from others. This work particularly deals with the trolls that spread misinformation, sow discord, and amplify existing conflicts to achieve political or ideological goals. A specific scenario where the described trolling behavior was seen is the Russian troll interference in the 2016 U.S. presidential elections, where they used social media platforms to disseminate divisive content and manipulate public opinion. Through this work, it is identified that such social media interactions can be captured better when formulated as graphs and hence proposed a novel heterogeneous graph representation of the user and their interactions that are analyzed using graph neural networks, a deep learning approach to identify a troll based on the user and the contents of his post on social media platforms. We have proposed the use of graph attention networks for the same after comparing with other competing graph neural network models, achieving a precision of 90.76% and AUC of 85.90% on our test data.