Source-linked AI summary
Network-based Fake News Detection: A Pattern-driven Approach
Xinyi Zhou, Reza Zafarani
TL;DR
Fake news detection research has largely relied on news content, while network-based clues from propagation and spreaders have been less comprehensively explored. The paper proposes an explainable pattern-driven network approach, represents patterns across multiple network levels, and reports comparatively strong real-world performance, with stable early detection using limited propagation information.
Problem
Existing fake news detection approaches heavily rely on news content, while network-based clues from propagation and spreaders have been insufficiently explored comprehensively and explainably.
Method
The approach uses empirically and theoretically supported fake news dissemination patterns, represented as features at node, ego, triad, community, and network levels.
Results
Experiments on real-world data show comparatively good performance against state-of-the-art methods, with stable performance using limited propagation information.
Takeaways & Limitations
The approach provides an explainable, relatively content-independent perspective for detecting fake news and supports early detection from limited network information.
Takeaways & Limitations
Compared with content-based models, the approach can hardly detect fake news before propagation and still requires validation across domains and languages.
Abstract
from arXiv · showhide
Fake news gains has gained significant momentum, strongly motivating the need for fake news research. Many fake news detection approaches have thus been proposed, where most of them heavily rely on news content. However, network-based clues revealed when analyzing news propagation on social networks is an information that has hardly been comprehensively explored or used for fake news detection. We bridge this gap by proposing a network-based pattern-driven fake news detection approach. We aim to study the patterns of fake news in social networks, which refer to the news being spread, spreaders of the news and relationships among the spreaders. Empirical evidence and interpretations on the existence of such patterns are provided based on social psychological theories. These patterns are then represented at various network levels (i.e., node-level, ego-level, triad-level, community-level and the overall network) for being further utilized to detect fake news. The proposed approach enhances the explainability in fake news feature engineering. Experiments conducted on real-world data demonstrate that the proposed approach can outperform the state of the arts.
1. INTRODUCTION
The paper motivates automatic fake news detection by identifying limits of manual fact-checking and content-focused methods. It proposes an explainable, network-based pattern-driven approach that represents dissemination patterns across multiple network levels and performs well on real-world data.
- Manual fact-checking does not scale with the volume of newly created social-media information, motivating automatic fake news detection.
- The proposed model uses news-dissemination patterns involving spread distance, spreader participation, engagement, and network connectedness.
- These patterns are quantified at node, ego, triad, community, and overall network levels for explainable supervised detection.
- Experiments on real-world data show comparatively good performance against state-of-the-art methods.
- Empirical studies and social psychological theories are used to validate and interpret fake news patterns concerning news, spreaders, and their relationships.
2. RELATED WORK
Related work divides fake news detection into content-based and network-based approaches. Network-based methods use social context from propagation and relationships among news, publishers, users, and posts, while content-focused methods can omit propagation information and be vulnerable to manipulated writing styles.
- Existing fake news detection research is generally grouped into content-based and network-based approaches.
- Content-based methods assess authenticity through news content, including knowledge extracted into subject-predicate-object triples and compared with ground truth.
- Content-focused approaches may omit auxiliary propagation information such as news spreaders and can be sensitive to manipulated writing styles.
- Network-based detection uses social context from homogeneous or heterogeneous networks representing stances, propagation trees, and relationships among news, publishers, users, and posts.
3. FAKE NEWS PATTERNS AND REPRESENTATION IN NETWORKS
The paper identifies distinguishable fake-news dissemination patterns and represents them as explainable features across multiple network levels. These features describe spreaders, their relationships, network structure, and how far or strongly news propagates.
- Fake news dissemination is represented across node, triad, community, and overall-network levels for supervised detection.The representation covers patterns involving news, spreaders, and relationships among spreaders.
- More-Spreader Pattern: More users spread fake news than true news, a pattern quantified through general and specific spreader features.Features include spreader counts and proportions, alongside susceptibility and influence characteristics.
- More-Spreader Pattern: User susceptibility is measured from fake-news involvement counts or spreading frequency, then used to label users with a threshold.Susceptibility scores range conceptually from users spreading only true news to users spreading only fake news.
- Farther-Distance Pattern: Fake news spreads farther than true news, represented by network diameters computed with geodesic and effective distances.Effective distance incorporates information flow through weighted networks, while geodesic distance uses shortest paths.
- Stronger-Engagement and Denser-Network Patterns: Spreaders engage more strongly with fake news, while fake-news spreaders also form denser networks than truth spreaders.The integrated representation combines network diameter, size, and density to capture overall network structure.
4. EXPERIMENTS
The experiments verify the proposed fake news detection approach and evaluate its effectiveness through an experimental setup and results analysis.
- The experiments verify the effectiveness of the proposed fake news detection approach.
- The experimental procedure begins by presenting the setup before evaluating the experimental results.
- The setup covers the data, data preparation, and baseline methods used for comparison.
4.1 Experimental Setup
The evaluation uses two public benchmark datasets with expert-provided labels and Twitter-network information, then compares the proposed method with content-, network-, and hybrid-based baselines.
- 4.1.1 Data: Two public benchmark datasets use articles from PolitiFact and BuzzFeed, with true-or-fake labels provided by fact-checking experts.
- 4.1.1 Data: Both datasets include Twitter users and their following relationships alongside news content and labels.
- 4.1.2 Data Preparation: Feature values are computed for supervised learning, while user susceptibility scores use users’ historical fake-news dissemination behavior.
- 4.1.3 Baselines: The proposed method is compared on the same datasets with content-based, network-based, and hybrid fake news detection methods.
- 4.1.3 Baselines: Content-based baselines include linguistic features such as n-grams, CFGs, LIWC proportions, readability, and latent Word2Vec or Doc2Vec representations.
- 4.1.3 Baselines: Network and hybrid baselines use information from user profiles, tweets, propagation trees, or embedded relationships among news, publishers, and spreaders.
4.2 Performance Evaluation
The evaluation tests classifier, pattern, feature, susceptibility, and data-availability effects, finding strong performance across datasets and robustness under varying conditions.
- General Performance Evaluation: Random forests perform best, achieving accuracy and F1 scores around 0.93 on PolitiFact and around 0.84 on BuzzFeed.The experiments compare SVM, k-NN, NB, DT, and RF classifiers.
- Performance of Fake News Patterns: More-Spreader and Stronger-Engagement Patterns perform best individually, while combining patterns generally improves performance over separate patterns and network similarity.Combined patterns achieve accuracy and F1 scores around 93% on PolitiFact and 82% on BuzzFeed.
- General Performance Evaluation: The proposed network-based approach performs relatively well against content-based and network-based baselines and is comparable to a hybrid baseline.Adding linguistic features enables it to outperform the cited hybrid baseline.
- Performance of Fake News Patterns: Network similarity slightly improves the four-pattern combination, reaching accuracy and F1 scores around 93% on PolitiFact and 84% on BuzzFeed.The combined patterns perform significantly better than network similarity alone.
- Sensitivity Analysis: User-susceptibility features improve fake-news prediction by approximately 10%, while the two susceptibility-computation methods perform similarly.The methods use either the number or frequency of fake-news stories spread by a user.
- Impact of News Number and Distribution: Across varying news quantities and label distributions, the approach generally achieves accuracy of approximately 0.7–0.85 and F1 scores of approximately 0.65–0.9.The evaluation varies both the number and distribution of available news articles.
5. CONCLUSION
The approach uses explainable fake-news patterns across multiple social-network levels and performs relatively well against the state of the art on two real-world datasets. Its main boundary is that, unlike content-based models, it generally requires some propagation before detection.
- The approach represents fake-news patterns at node, ego, triad, community, and overall network levels for explainable detection.
- Experiments on two real-world datasets validate the approach's effectiveness and show relatively strong state-of-the-art performance.
- Stable performance with limited propagation information and few training articles supports the approach's suitability for early detection.
- Compared with content-based models, the approach can hardly detect fake news before it has propagated on social media.