TopoMSG: A Topology-Aware Multi-Scale Graph Network for Social Bot Detection

ACL ARR 2025 May Submission7688 Authors

Abstract: Social bots are automated programs designed to spread rumors and misinformation, posing significant threats to the security of the network. Graph Neural Network (GNN)-based social bot detection models are limited by the over-smoothing and over-squashing problems of the message-passing mechanism, making it difficult to effectively extract key high-dimensional topological features and model complex topological structures across different social networks. To address the issue of limited topological feature extraction caused by over-smoothing and over-squashing in GNN-based social bot detection models, we propose a topology-aware multi-scale detection method for social bots. By leveraging local topological layers and a clustering attention mechanism, the approach effectively incorporates topological features into node representations and captures multi-level structural patterns at both global and local scales. Experimental results demonstrate that our model exhibits strong competitiveness on three widely used benchmark datasets, effectively addressing existing methods' limitations in capturing local feature patterns, while also being capable of capturing global features, thereby enhancing the overall modeling of complex structures. We publicly release our code in https://anonymous.4open.science/r/TopoMSG-2C41/

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: Social Bot Detection,GNN,Topological Data Analysis

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 7688