.png){kind=icon}
.png){kind=icon}
Return
SeBot-MLS: Multi-level structural feature learning for graph-based social bot detection
.png){kind=icon}
OA
AI
DOI：10.1016/j.jnlssr.2026.100310.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recent advances in artificial intelligence technologies, especially large language models, have made social bots more intelligent and human-like. This increased sophistication poses significant challenges to accurate bot detection. Currently, most existing studies leverage graph neural networks, which employ propagation and aggregation mechanisms to learn user profiles and textual features. However, these approaches inadequately exploit explicit structural features and lack a systematic methodology for their extraction. To address these limitations, this paper proposes SeBot-MLS, a multi-level social bot detection framework. SeBot-MLS extends a structural entropy guided social bot detector (SeBot) through a dual-branch architecture: it retains the original multi-view user-level feature extraction branch of SeBot while introducing a parallel branch for multi-level structural feature encoding and fusion (MLS). This new branch captures four levels of structural features: dyadic relationship, polyadic relationship, local structure, and global structure. The features are subsequently encoded through a hierarchical graph neural network for layered representation learning, then fused across levels using adaptive weights and multi-head self-attention. Experimental results on the MGTAB and Twibot-20 datasets demonstrate that SeBot-MLS significantly improves social bot detection performance compared with SOTA methods.
Keywords:
Social bot detection
Graph neural networks
Multi-level structural features
Attention mechanisms
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
3.4
Papers:
290
Citations:
560