SAMGAT: structure-aware multilevel graph attention networks for automatic rumor detection.

Area of Science:

• Computer Science
• Social Computing
• Artificial Intelligence

Background:

• The rapid spread of unverified information on social media poses significant societal risks.
• Existing rumor detection methods struggle to capture propagation structures and filter irrelevant comments.

Purpose of the Study:

• To introduce a novel approach, the Structure-Aware Multilevel Graph Attention Network (SAMGAT), for effective rumor classification.
• To enhance the accuracy and early detection capabilities of fake news identification systems.

Main Methods:

• Developed SAMGAT, a graph attention network incorporating dynamic attention (GATv2 and dot-product) to model contextual relationships.
• Implemented a structure-aware attention mechanism to reflect rumor propagation structures by learning edge-indicating attention weights.
• Utilized top-k attention filtering and claim-guided attention pooling to focus on critical structural features and informative posts.

Main Results:

• SAMGAT demonstrated superior performance compared to state-of-the-art methods on benchmark datasets for rumor detection.
• The proposed model significantly improved the effectiveness of early rumor detection.

Conclusions:

• SAMGAT offers a robust framework for analyzing complex information propagation patterns on social media.
• This approach enhances the ability to identify and mitigate the spread of misinformation.