MB-HGTBot: Modeling Multitype Social Behaviors in Heterogeneous Graphs for Robust Twitter Bot Detection

Abstract

Existing studies predominantly utilize the topology of the Twittersphere to counteract the challenge of social bots circumventing detection via feature manipulation. However, the heterogeneity of users’ social behavioral patterns remains largely underexplored. To address this issue, we propose MB-HGTBot, a social behavior scenario-oriented framework for Twitter bot detection. Our approach effectively extracts and integrates multimodal user semantic representations with numerical attributes while optimizing feature learning to improve the accuracy of bot detection. Specifically, we construct heterogeneous information networks (HINs) based on diverse social behaviors, enabling a more comprehensive characterization of structural variations underlying diverse social interactions, thereby significantly improving detection performance. Experimental results show that MB-HGTBot outperforms state-of-the-art methods, achieving an accuracy of 88.29% across six distinct social behavior patterns in the Twibot-22 dataset and three real-world topic-specific contexts. Furthermore, our findings highlight the pivotal role of social behavior diversity in Twitter bot detection and provide novel insights into addressing the challenges posed by manipulated features. The implementation of MB-HGTBot is available at https://github.com/Bemcliu/MB_HGTBot.