SEGCN: a subgraph encoding based graph convolutional network model for social bot detection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Network Science

Background:

• Graph Convolutional Networks (GCNs) are used for social bot detection by analyzing node features.
• The expressive power of standard GCNs is limited by the 1st-order Weisfeiler-Leman isomorphism test, hindering optimal bot detection.
• Existing GCN models struggle to capture complex relational patterns crucial for differentiating bots from genuine users.

Purpose of the Study:

• To propose a novel GCN model with enhanced expressive power for improved social bot detection.
• To address the limitations of current GCNs in capturing intricate network structures relevant to bot identification.
• To develop a more accurate and robust method for detecting social bots in online networks.

Main Methods:

• Developed a subgraph encoding based GCN model named SEGCN.
• SEGCN computes node representations by encoding surrounding induced subgraphs, not just immediate neighbors.
• The model's architecture enhances its expressive power beyond the 1st-order Weisfeiler-Leman test.

Main Results:

• SEGCN demonstrated improved performance on social bot detection tasks.
• The model achieved approximately 2.4% and 3.1% accuracy improvements on the Twibot-20 and Twibot-22 datasets, respectively.
• Experimental results confirm SEGCN's superiority over state-of-the-art social bot detection models.

Conclusions:

• The proposed SEGCN model offers a significant advancement in social bot detection capabilities.
• Subgraph encoding provides a more powerful approach to node representation learning in GCNs for this task.
• SEGCN represents a more effective solution for identifying malicious social bots in online environments.