Multirelational Twitter Bot Detection Using Graph Neural Networks

Social media is an important means of modern communication and information exchange. Ease of access and global reach are the primary factors contributing to the popularity of several social media platforms. Apart from human (real) users, they also have bots, automated programs developed to facilitate online engagement and interaction. However, these bots are increasingly being exploited for malicious purposes, such as disseminating false information and manipulating public opinion. Detecting and removing bots is essential for maintaining authentic human interactions and trustworthy public opinions. This paper proposes the Multirelational Bot Detection Graph Neural Network method (MultiBotGNNs) to detect bots. It treats the problem as a node classification problem in a graph that has two types of nodes (real and bot users) and multiple types of edges, e.g., the various relations (interactions) between users. To detect bots, we use their characteristics processed along the social network of the different user relations, such as follow, follower, and like. We experimented with the TwiBot22 dataset, which is a big dataset of bots on X (formerly Twitter). Our hypothesis is that by considering just the different relations and their characteristics for the prediction, and by ignoring the tweet contents, we can get good results without the extra process time for the tweets. Experimental results confirm our hypothesis.