Contrastive Learning-Driven Fake News Detection: Preserving Semantics, Unveiling Distortions

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Information Science

Background:

• Fake news detection is challenged by data scarcity and semantic integrity loss.
• Existing methods struggle with real-world social network data complexity and augmentation limitations.

Purpose of the Study:

• To propose a contrastive learning-driven fake news detection (CLFD) framework.
• To address limitations of existing methods in data scarcity and semantic integrity destruction.
• To develop a universal and portable fake news detection method using only textual content.

Main Methods:

• Developed a CLFD framework utilizing distortion-reversion dual-view manipulation.
• Employed learnable neural networks to simulate nonlinear information transformations for contrastive view generation.
• Implemented distortion-aware contrastive learning and multiobjective joint optimization strategies.

Main Results:

• CLFD effectively generates diverse views while preserving semantic integrity, solving semantic destruction issues.
• The framework achieves efficient detection using only textual content, without requiring propagation structures.
• Demonstrated superior performance in accuracy, robustness, and generalization across benchmark datasets.

Conclusions:

• The proposed CLFD framework offers a significant advancement in fake news detection.
• The method exhibits high universality and portability due to its reliance solely on textual content.
• CLFD enhances the capability to capture deceptive features, outperforming state-of-the-art methods.