[Paper Review] GAMC: An Unsupervised Method for Fake News Detection using Graph Autoencoder with Masking

Shu Yin, Chao Gao+1 more|arXiv (Cornell University)|Dec 10, 2023

Misinformation and Its Impacts4 citations

TL;DR

GAMC proposes an unsupervised fake news detection method using a graph autoencoder with masking and contrastive learning, leveraging self-supervised signals from data augmentation (node masking and edge dropping) to eliminate reliance on labeled data. It achieves state-of-the-art performance on PolitiFact (82.8% accuracy, 82.3% F1) and GossipCop (94.1% accuracy, 93.7% F1), demonstrating robustness without manual annotation.

ABSTRACT

With the rise of social media, the spread of fake news has become a significant concern, potentially misleading public perceptions and impacting social stability. Although deep learning methods like CNNs, RNNs, and Transformer-based models like BERT have enhanced fake news detection, they primarily focus on content, overlooking social context during news propagation. Graph-based techniques have incorporated this social context but are limited by the need for large labeled datasets. Addressing these challenges, this paper introduces GAMC, an unsupervised fake news detection technique using the Graph Autoencoder with Masking and Contrastive learning. By leveraging both the context and content of news propagation as self-supervised signals, our method negates the requirement for labeled datasets. We augment the original news propagation graph, encode these with a graph encoder, and employ a graph decoder for reconstruction. A unique composite loss function, including reconstruction error and contrast loss, is designed. The method's contributions are: introducing self-supervised learning to fake news detection, proposing a graph autoencoder integrating two distinct losses, and validating our approach's efficacy through real-world dataset experiments.

Motivation & Objective

• To address the challenge of fake news detection in social media without relying on costly, time-consuming labeled datasets.
• To incorporate social context and propagation dynamics into fake news detection by modeling news as a graph.
• To develop a self-supervised learning framework that leverages both content and structural features of news propagation graphs.
• To eliminate the need for manual annotation by using data augmentation as a self-supervised signal.
• To improve detection performance through a composite loss function combining reconstruction and contrastive learning.

Proposed method

• The method performs data augmentation on the original news propagation graph via random node feature masking and edge dropping to generate two enhanced views.
• A graph encoder processes the augmented graphs to produce graph-level representation vectors that capture global structure and propagation dynamics.
• A graph decoder reconstructs the original graph from the representation vectors, minimizing reconstruction error to preserve structural and semantic patterns.
• A composite loss function combines reconstruction loss (L2 distance between original and reconstructed graphs) and contrastive loss (minimizing distance between representations of augmented views from the same graph).
• The model is trained end-to-end using the composite loss, enabling self-supervised representation learning without labels.
• The final graph-level representations are used directly for fake news classification after training.

Experimental results

Research questions

• RQ1 Can self-supervised learning effectively replace supervised signal in fake news detection without requiring labeled data?
• RQ2 How does incorporating both content and propagation context improve fake news detection beyond text-only models?
• RQ3 What is the impact of data augmentation (masking and edge dropping) on the model’s ability to learn robust representations?
• RQ4 How do reconstruction and contrastive losses jointly contribute to improved detection performance?
• RQ5 Can a graph autoencoder with masking and contrastive learning achieve competitive results compared to supervised baselines on real-world datasets?

Key findings

• GAMC achieved 82.8% accuracy and 82.3% F1 on the PolitiFact dataset, outperforming all compared supervised methods including BiGCN and GACL.
• On the GossipCop dataset, GAMC achieved 94.1% accuracy and 93.7% F1, significantly outperforming the next best method (GACL) by 3.4 percentage points in accuracy.
• Ablation studies confirmed that removing data augmentation (GAMC-Aug) reduced accuracy to 79.3% on PolitiFact, proving its critical role in representation learning.
• Removing the reconstruction loss (GAMC-Lrec) led to a drop in F1 to 75.8% on PolitiFact, indicating its importance in preserving structural fidelity.
• The contrastive loss component (GAMC-Lcon) was essential, as using only one augmented view reduced performance, showing that contrastive learning enhances discrimination.
• Optimal hyperparameters were found at a mask rate of 0.5 and edge drop rate of 0.2, where model performance peaked on both datasets.