Earticle

초록

영어

Social bot detection on emerging platforms is constrained by sparse labels and by evaluation practices that focus on classification accuracy while ignoring network validity for network analytics. This study proposes a graph-based semi-supervised bot detection framework for Bluesky. From 2.2 million accounts, we construct 6,891 high-confidence bot seeds and 683,209 conservative non-bot anchors without treating unlabeled accounts as humans. We then introduce a Role-Discrepancy GNN that models directional role asymmetry across repost, reply, quote, and mention relations. On a temporal test split, the proposed model achieves the best overall performance among profile-only and graph-based baselines, with AUC = 0.9930 and F1 = 0.8936. Network validation further shows the strongest structural separation, with modularity Q = 0.0547 and mean homophily gap ΔH = 0.0515. These results support relation-role discrepancy as a useful signal for bot detection under sparse labels.

목차

Abstract
Introduction
Related Works
Social Network Analysis
Social Bots
Social Bot Detection
Graph-based Bot Detection Models
Methods
Research Design
Data
Proposed Model
Evaluation
Results
Classification
Network Separation
Conclusions
Discussions
Limitations
Future Works
Acknowledgments
References

키워드

저자

• 현상훈 [ Department of Business Informatics, Hanyang University ]
• 신민수 [ Department of Management Information Systems, Business School, Hanyang University ] Corresponding Author

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  한국경영정보학회 정기 학술대회 [KMIS Conference]

• 간기

  반년간

• 수록기간

  1990~2026

• 십진분류

  KDC 325 DDC 658