KAGN:knowledge-powered attention and graph convolutional networks for social media rumor detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Online and social media platforms generate numerous rumor posts, necessitating effective automatic detection methods.
• Existing rumor detection models primarily focus on linguistic and semantic content, often overlooking valuable external knowledge.
• Integrating knowledge entities and concepts can significantly enhance the accuracy of rumor detection systems.

Purpose of the Study:

• To propose a novel end-to-end attention and graph-based neural network model (KAGN) for rumor detection.
• To incorporate external knowledge from knowledge graphs to address limitations in current methods.
• To improve the accuracy and robustness of automatic rumor detection by leveraging both semantic and knowledge-level information.

Main Methods:

• Developed a knowledge-aware attention mechanism to fuse local knowledge from identified entities and concepts.
• Constructed a graph integrating post texts, entities, and concepts for analysis.
• Utilized graph convolutional networks (GCNs) to explore long-range knowledge dependencies within the graph structure.
• Proposed a novel end-to-end attention and graph-based neural network model (KAGN).

Main Results:

• The KAGN model demonstrated superior or comparable performance against state-of-the-art methods on four real-world datasets.
• Experiments validated the significant contribution of external knowledge integration to rumor detection.
• The model effectively combines semantic-level and knowledge-level representations for enhanced detection.

Conclusions:

• The proposed KAGN model offers an effective approach to automatic rumor detection by integrating external knowledge.
• Leveraging knowledge graphs alongside textual content significantly improves rumor detection capabilities.
• KAGN represents a promising advancement in combating the spread of misinformation online.