计算机科学
图形
图论
人工智能
理论计算机科学
计算机网络
算法设计
电子邮件
社会网络分析
社交网络(社会语言学)
网络拓扑
数据挖掘
智能网
信号处理
作者
Zekun Liu,Yaguang Lin,Xiaoming Wang
标识
DOI:10.1109/tifs.2026.3744622
摘要
Social bots are automated programs closely mimicking human behavior, frequently employed to manipulate public opinion and disseminate false information. The ultimate objective of social bot detection is to identify and differentiate social bots from genuine users through intelligent algorithms, thereby preserving the authenticity and security of the online environment. Although graph neural network-based detection methods can effectively aggregate node features via message passing mechanisms, existing approaches often succumb to adversarial camouflage where bots mimic human behaviors to evade topological detection, and struggle to reconcile the structural mismatch between numerical metadata and high-dimensional text. Furthermore, traditional GNNs fail to account for the inherent unreliability of interaction edges maliciously injected by bots to pollute message-passing mechanisms. To address these challenges, this paper proposes HEDGE, a Hyperbolic-Enhanced and Dynamic Edge-Aware Graph Network, for social bot detection. First, we introduce Hyperbolic Space Linear Transformation (HSLT) to model user hierarchical features and mitigate the asymmetric interference between meta-attributes (e.g., numerical and boolean attributes) and text features. Second, we design a variational graph auto-encoder and edge feature optimizer to achieve discriminative feature fusion through probabilistic edge feature modeling and attention mechanisms. Finally, we construct a dynamic graph network with unreliable edge-awareness to adaptively aggregate multi-relational features based on reliability scores while suppressing the effects of spurious interactions. Experimental results show that HEDGE outperforms 14 baselines on three real-world datasets. Furthermore, the HSLT module effectively captures hierarchical differences between nodes and exhibits architectural universality.
科研通智能强力驱动
Strongly Powered by AbleSci AI