Beyond Following: Augmenting Bot Detection with the Integration of Behavioral Patterns

Social media platforms like Twitter revolutionized online communication. But this new era of interaction has brought with it a challenge—the widespread presence and influence of bot accounts. These bots are rapidly evolving, making traditional detection methods increasingly ineffective and allowing malicious actors to influence public discourse. While existing bot detection methods report high performance, such results might actually be connected to shortcomings in dataset collection and labeling practices, rather than reflecting their true ability to detect bots, casting doubt on their true reliability. Our study introduces higher-order behavior-based relations, including Co-Retweet, and Co-Hashtag, derived from the TwiBot-22 dataset. By leveraging these new relations in the BotRGCN architecture, we shift the emphasis from isolated accounts to coordinated group dynamics, making it more challenging for bot developers to evade detection. This strategy not only acknowledges the limitations and inherent biases presented in existing bot detection techniques, but also presents a way to address them. Our experiments support this approach as a promising way forward to tackle challenges in bot detection.