Abstract
Fake news detection is a complex challenge involving both linguistic content and social propagation patterns. This paper introduces FND-NUP, a multi-contextual framework that integrates news content, user profiles, and propagation networks. By employing a co-attention mechanism, the model captures the interplay between textual features and social context, achieving state-of-the-art performance on benchmark datasets while providing interpretable explanations for its predictions.
Introduction
Fake news is not merely a content problem; it is a social phenomenon. While traditional models often focus exclusively on linguistic cues or propagation patterns, both approaches possess inherent blind spots. Fake news is frequently authored to mimic the style of legitimate journalism, complicating text-only detection. Conversely, models focusing solely on "viral" paths may overlook the underlying reasons for a story's spread.
The authors identify a critical gap in "multi-contextual" detection: the necessity to fuse news content, user credibility, and the structural evolution of propagation networks into a single, interpretable model.
Methodology
The proposed FND-NUP framework (Fake News Detection with News content, User profiles, and Propagation networks) operates in three primary stages:
- Embedding Learning:
- Users: The model utilizes Graph Attention Networks (GAT) to learn representations of users based on profile features (e.g., verified status, follower count) and their recent 200 tweets.
- Content: News articles are encoded using pre-trained word embeddings such as word2vec or BERT.
- Co-Attention Mechanism: This layer serves as the core innovation. Rather than simply stacking features, the co-attention layer learns the relationship between the news text and user nodes. It identifies which specific words in the article and which users in the network are most indicative of "fakeness."
- Classification: The fused representation is processed through a feed-forward neural network to produce a binary classification (Fake or Real).
Experimental Setup
The study utilizes two widely recognized benchmarks from the FakeNewsNet repository:
- Politifact: Focuses on political news and fact-checking.
- Gossipcop: Focuses on entertainment and celebrity news. These datasets are particularly valuable as they include not only the articles but also the comprehensive list of users who shared them, enabling the construction of detailed propagation graphs.
Key Insights and Results
The "Echo Chamber" Signature
Research findings indicate that fake news propagation networks often form smaller, isolated clusters with reduced connectivity to the original source. These "echo chambers" serve as structural red flags that the model successfully identifies.
Credibility vs. Content
By integrating user profiles, the model weighs the credibility of spreaders. A story shared by numerous unverified accounts with low follower counts is treated differently than one shared by established entities.
Performance Metrics
FND-NUP consistently outperformed state-of-the-art baselines like BiGCN and UPFD:
- Gossipcop: Achieved an Accuracy of 0.965 and a Macro-F1 of 0.963.
- Politifact: Achieved an Accuracy of 0.869 and a Macro-F1 of 0.865.
Interpretability
Because the model uses attention weights, it can generate heatmaps pointing to sensationalist words (e.g., "exclusively," "nightmare") and specific low-credibility edges in the social graph to justify its predictions. The average attention score for real news networks (0.0244) was nearly double that of fake news networks (0.0125), confirming that real news spreaders tend to have stronger, more credible correlations.
Conclusion and Future Work
While FND-NUP demonstrates significant power, several areas for future development remain. Constructing propagation networks is currently hindered by social media API constraints, which often require researchers to make assumptions about retweet paths. Future work will explore the integration of more advanced Graph Neural Networks (GNNs) and Natural Language Processing (NLP) models to further enhance performance and deepen the sophistication of the "why" behind model explanations.
Summary updated