Graph-Based Temporal Behavior Analysis for Early Detection of Coordinated Malicious Accounts in Social Media Platforms

Abstract

The proliferation of coordinated malicious accounts poses significant threats to social media platform integrity and online discourse quality. This research proposes a comprehensive detection framework integrating heterogeneous graph neural networks with temporal behavior analysis to identify coordinated account clusters before large-scale malicious activities manifest. Our approach constructs multi-relational social graphs capturing follower networks, retweet cascades, and mention patterns while extracting time-series behavioral features including posting frequency distributions, coordination windows, and synchronized activity signatures. Experimental validation on real-world Twitter datasets demonstrates that the proposed framework achieves 89.7% detection accuracy with 87.3% F1-score, outperforming baseline methods by 4.3-17.2% across different comparison approaches. Ablation studies reveal that temporal coordination features contribute 6.7 percentage points performance improvement while heterogeneous graph structures provide 5.2 percentage points accuracy gains. The framework enables early warning capabilities detecting coordinated campaigns 4.7 days before peak malicious activity deployment.