BotSCL: Heterophily-Aware Social Bot Detection with Supervised Contrastive Learning

Qi Wu, Yingguang Yang, Buyun He, Hao Liu, Renyu Yang, Yong Liao

Abstract: Detecting social bots, which continuously evolve, presents an escalating challenge. Although graph-based detection techniques utilize various relationships within social networks to model node behavior, they often fail to account for inherent heterophily–connections between different types of accounts. When message passing occurs across heterophilous edges, it can cause feature blending between bots and legitimate users, leading to indistinct representations. To address this issue, we propose BotSCL, a contrastive learning framework that is aware of heterophily. BotSCL adapts by differentiating between representations of heterophilous neighbors while aligning representations of homophilous ones. Our approach employs two graph augmentation strategies to create varied graph views and introduces a channel-wise, attention-free encoder to address the limitations of traditional neighbor information aggregation. Supervised contrastive learning then helps the encoder focus on aggregating information specific to each class. Extensive experiments on two real-world social bot detection datasets reveal that BotSCL outperforms existing baseline models, including advanced bot detection methods, as well as techniques based on partial heterophily and graph contrastive learning.