Abstract
Traditional automated fake news detection often relies solely on the textual content of individual articles, ignoring the broader context of the publisher and the circumstances of the news release. This paper introduces the Publisher-oriented Multi-view Graph Model (PMGM), a novel framework that integrates publisher metadata, historical reliability, and linguistic style. By constructing multi-view graphs that capture relationships based on topics and social occasions, PMGM significantly improves detection accuracy. Evaluation on the LIAR and Weibo21 datasets demonstrates that PMGM outperforms existing state-of-the-art methods, particularly in complex multi-class classification tasks.
Introduction
Most current fake news detectors suffer from "tunnel vision," focusing almost exclusively on the content of a single article. However, fake news does not exist in isolation. A publisher's historical reliability is a strong indicator of the veracity of their future output. Current research often fails to utilize publisher metadata—such as political affiliation and past accuracy—and ignores the nuanced linguistic "style" that distinguishes deceptive text. Furthermore, existing graph-based models typically use simplistic relationships, failing to capture how news items are interconnected through specific social contexts or shared publishers.
Methodology: The PMGM Framework
The PMGM framework addresses these gaps through a three-stage process: enhanced representation, multi-view graph modeling, and graph convolution with attention-based fusion.
1. Enhanced News Representation
The model generates a comprehensive feature set for each news item:
- Semantic Encoding: BERT is utilized to capture the deep semantic meaning of the text.
- Publisher Profile: This includes the publisher’s identity, political party, and historical "credit score" derived from PolitiFact ratings (ranging from "true" to "pants-on-fire").
- Linguistic Style: The model extracts 40 quantifiable features, such as subjectivity, sentiment polarity, readability scores, and "non-immediacy" (e.g., the frequency of first-person pronouns).
2. Multi-view Graph Modeling
The core innovation of PMGM is the construction of three distinct graph "views" to model inter-news relationships:
- Full-Connected View: Links all news items originating from the same publisher.
- Topic-Oriented View: Connects news from the same publisher that shares a common subject (e.g., healthcare or economy).
- Occasion-Oriented View: Connects news shared during similar types of events, such as televised interviews versus prepared press releases.
3. Graph Convolution and Fusion
These graphs are processed using Graph Convolutional Networks (GCNs) integrated with Jumping Knowledge Networks (JK-Nets). This setup allows the model to aggregate local and global network information effectively. An attention mechanism then weighs the importance of each view before a final classifier determines the news's veracity.
Experimental Results
The model was evaluated on two major benchmarks: the LIAR dataset (12,800 English news items with six truth labels) and the Weibo21 dataset (9,000 Chinese news items for binary classification).
Performance Gains
PMGM consistently outperformed state-of-the-art baselines like M-GCN and MMFake. On the notoriously difficult LIAR dataset, PMGM achieved an accuracy of 54.3% and a Macro-F1 of 58.2%, a significant improvement over previous models which typically range between 44% and 47%. On the Weibo21 dataset, the model reached 89.4% accuracy, proving its effectiveness across different languages and social media contexts.
Key Insights
- The Power of Credit History: Ablation studies showed that removing publisher profiles led to a 25% drop in performance, confirming that a publisher's track record is a primary predictor of truthfulness.
- Contextual Reliability: The "Occasion-Oriented" graph revealed that publishers' accuracy varies by context; for instance, a politician might be more accurate in a prepared statement than in a live interview.
- Style as a Signal: High subjectivity (scores above 0.7) was found to be a strong "red flag" for fake news, capturing linguistic patterns that standard semantic encoders might overlook.
Conclusion and Future Work
PMGM demonstrates that context—specifically publisher identity and the social occasion of news release—is vital for accurate fake news detection. While highly effective, the authors note that future work should incorporate temporal data (how news spreads over time) and audience interaction patterns to further enhance the model's robustness.
Summary updated