CB-MTE: Social Bot Detection via Multi-Source Heterogeneous Feature Fusion.

Area of Science:

• Artificial Intelligence
• Computer Science
• Social Network Analysis

Background:

• Social bots pose a significant threat by engaging in large-scale influence campaigns.
• Existing bot detection methods struggle with incomplete features and dynamic bot behaviors.
• The challenge lies in capturing bots' evolving tactics and collaborative nature.

Purpose of the Study:

• To propose a novel framework, CB-MTE, for enhanced social bot detection.
• To address the limitations of single-source feature-based detection methods.
• To improve the recognition of dynamic behavioral traits and collaborative bot activities.

Main Methods:

• CB-MTE utilizes multi-source heterogeneous feature fusion in a hierarchical architecture.
• Deep semantic representations from text (DistilBERT) and community-aware graph embeddings are employed.
• Manifold learning for dimensionality reduction and a CatBoost-based mechanism for collaborative reasoning are applied.

Main Results:

• CB-MTE significantly outperforms mainstream detection models on the TwiBot-22 dataset.
• The framework effectively captures dynamic behavioral traits of social bots.
• Collaborative bot activities are detected with higher accuracy.

Conclusions:

• CB-MTE demonstrates superior performance in social bot detection.
• Multi-source feature integration is crucial for capturing comprehensive bot characteristics.
• The proposed framework offers a robust solution for identifying sophisticated social bot networks.