Hybrid Deep Learning Model for Fake News Detection in Social Networks (Student Abstract)

Bibek Upadhayay, Vahid BehzadanAAAI 2022 · DOI

Abstract

Detecting fake news in social networks has become increasingly difficult as creators learn to mimic the writing styles of legitimate news outlets. This paper presents a hybrid deep learning architecture that combines textual analysis with social network propagation patterns. By integrating BERT-based sentiment analysis with Graph Attention Networks (GATs), the proposed model achieves up to 97% accuracy on benchmark datasets, demonstrating that structural dissemination patterns are a more robust indicator of news veracity than text alone.

Introduction

Automated fake news detection has traditionally relied on semantic analysis—identifying linguistic patterns or keywords associated with misinformation. However, as the "arms race" between detectors and creators intensifies, fake news has evolved to adopt professional journalistic styles, rendering text-only models insufficient. To address this, researchers are shifting focus toward the dissemination network. Fake news typically exhibits unique "cascade" patterns, traveling faster and reaching different audience depths than genuine news. This study proposes a hybrid model designed to fuse unstructured textual data with complex social graph structures to improve detection reliability.

Methodology

The proposed architecture processes information through two specialized branches before merging them for a final prediction.

Textual and Sentiment Branch

The model utilizes a BERT-base architecture combined with a Convolutional Neural Network (CNN) to extract high-level representations from news articles. Beyond raw text, the model incorporates affective features—emotions and sentiments—leveraging the Sentimental-LIAR dataset. This approach is based on the theory that fabricated statements often carry distinct emotional signatures designed to provoke specific reader reactions.

Network and Fusion Branch

To capture the social context, the model constructs a heterogeneous network representing relationships between users and the news they share. User-centric features, such as closeness centrality and historical sharing behavior, are extracted. This heterogeneous graph is converted into a homogeneous network and processed by a Graph Attention Network (GAT).

Figure 1: Architecture of the hybrid deep learning model showing the fusion of textual, sentimental, and social network features.

The GAT employs a self-attention mechanism to weigh the importance of different users within the network. The resulting "news-node embeddings" represent the collective social context of a story, which is then passed through a final CNN classifier to produce the veracity prediction.

Experimental Results

The model was evaluated using the FakeNewsNet benchmark, specifically the Politifact and BuzzFeed datasets. Both datasets provide a balanced mix of real and fake news stories along with their associated social network metadata.

  • Politifact: Achieved 97% accuracy, a significant improvement over the ~70% accuracy typical of style-only models.
  • BuzzFeed: Achieved 89.50% accuracy, confirming the model's ability to generalize across different social contexts.

Convergence and Training Dynamics

Analysis of the training process revealed that while the model converges effectively, the training loss exhibits noticeable fluctuations. This "noisy" behavior is likely due to the high parameter count of the deep learning architecture relative to the modest number of labeled news samples (e.g., 240 items in Politifact).

Figure 2: Comparison of training and validation loss over 300 epochs, illustrating model convergence and training noise.

Discussion and Conclusion

The results validate the hypothesis that social network structure is a powerful signal for fake news detection. While bad actors can mimic professional writing, they struggle to replicate the organic, multi-layered spread of legitimate news.

However, limitations remain. The small size of current labeled news datasets makes the model prone to local convergence issues. Future work will focus on scaling this architecture for real-time detection on massive, dynamic platforms like Facebook and Reddit, where computational overhead and evolving network structures present additional challenges.

Summary updated

BibTeX
@inproceedings{upadhayay2022hybrid,
  author = {Bibek Upadhayay and Vahid Behzadan},
  title = {{Hybrid Deep Learning Model for Fake News Detection in Social Networks (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21670},
}
Figure