Source-linked AI summary
Beyond News Contents: The Role of Social Context for Fake News Detection
Kai Shu, Suhang Wang, Huan Liu
TL;DR
The paper addresses fake news detection beyond news content by exploiting social context among publishers, news pieces, and users. It proposes TriFN to model publisher-news relations and user-news interactions simultaneously, with experiments showing effective detection and promising early-stage performance.
Problem
Existing fake news detection methods primarily rely on news content, while social context among publishers, news pieces, and users offers complementary information for detection.
Method
TriFN jointly models publisher-news relations and user-news interactions, along with news contents and user social relationships, for fake news prediction.
Results
TriFN demonstrates effective fake news prediction on two real-world datasets and achieves average relative improvements of 4.72% and 5.84% on BuzzFeed, and 5.91% and 4.39% on PolitiFact, for Accuracy and F1 score against LIWC+Castillo.
Takeaways & Limitations
Modeling publisher-news and user-news tri-relationships supports fake news detection and enables promising performance at the early stage of news dissemination.
Takeaways & Limitations
The framework focuses on user-news interactions where users share news without comments, treating these users as having aligned viewpoints with the news items.
Abstract
from arXiv · showhide
Social media is becoming popular for news consumption due to its fast dissemination, easy access, and low cost. However, it also enables the wide propagation of fake news, i.e., news with intentionally false information. Detecting fake news is an important task, which not only ensures users to receive authentic information but also help maintain a trustworthy news ecosystem. The majority of existing detection algorithms focus on finding clues from news contents, which are generally not effective because fake news is often intentionally written to mislead users by mimicking true news. Therefore, we need to explore auxiliary information to improve detection. The social context during news dissemination process on social media forms the inherent tri-relationship, the relationship among publishers, news pieces, and users, which has potential to improve fake news detection. For example, partisan-biased publishers are more likely to publish fake news, and low-credible users are more likely to share fake news. In this paper, we study the novel problem of exploiting social context for fake news detection. We propose a tri-relationship embedding framework TriFN, which models publisher-news relations and user-news interactions simultaneously for fake news classification. We conduct experiments on two real-world datasets, which demonstrate that the proposed approach significantly outperforms other baseline methods for fake news detection.
1 INTRODUCTION
Social media expands access to news but also enables widespread fake-news propagation, making content-only detection insufficient. TriFN therefore exploits complementary publisher-news and user-news social context for fake-news detection.
- Fake news is intentionally written to mislead consumers, making detection based solely on news content nontrivial.
- Publisher partisan bias correlates with news veracity, with more partisan publishers described as more likely to distort facts and publish fake news.
- User credibility and social connections provide additional signals because less credible users are more likely to spread fake news and like-minded users tend to share confirmatory news.
- TriFN models publisher-news relations and user-news interactions simultaneously, using their complementary social-context perspectives to improve fake-news prediction.
2 PROBLEM STATEMENT
The problem is to predict labels for unlabeled news using content, social, publishing, partisan, and partially labeled data. The formulation represents user relationships, user-news sharing, and publisher-news publishing as matrices.
- The framework includes news contents embedding, user embedding, user-news interaction embedding, publisher-news relation embedding, and news classification.
- The user adjacency matrix A records whether pairs of users are friends, while W records whether users shared news pieces.
- The publisher-news publishing matrix B represents which publishers published which news pieces.
- The task predicts the remaining unlabeled news label vector yU from labeled news and article, user, publishing, and publisher-partisan information.
3 A TRI-RELATIONSHIP EMBEDDING FRAMEWORK
TriFN models fake-news signals across news contents, user social relations and credibility-weighted interactions, publisher relations, and semi-supervised labels in one framework.
- Framework components: TriFN combines five components: news-content, user, user-news interaction, publisher-news relation, and semi-supervised classification embeddings.The framework learns latent representations from content and social context before predicting fake-news labels.
- News contents embedding: NMF maps the news-word matrix X into nonnegative latent matrices D and V, representing news pieces and words in a low-dimensional space.The regularization term λ(∥D∥2 F ) is introduced to avoid over-fitting.
- User embedding: User embedding applies NMF to the user adjacency matrix A to learn nonnegative user representations U from social relationships.The method uses a user-user correlation matrix T and weighted observed and negatively sampled links.
- User-news interaction embedding: Credibility-weighted user-news embedding brings high-credibility users closer to true news and low-credibility users closer to fake news in latent space.Credibility scores are inferred from user-generated data, with larger ci indicating higher credibility.
- Publisher relation and classification: Publisher-news embedding regularizes news representations using publisher partisan labels, while semi-supervised classification maps latent news features to fake-news labels.Publisher features are represented through the features of the news each publisher has published.
4 AN OPTIMIZATION ALGORITHM
TriFN is optimized by alternating updates of its variables rather than joint optimization, then predicts unlabeled-news labels after convergence.
- Optimization strategy: Because joint optimization is nonconvex, TriFN uses alternating least squares to update variables separately.The objective is denoted L, and separate updating rules are introduced for each variable.
- Update rules: The optimization updates U, V, T, p, q, and D using partial derivatives, least-squares solutions, nonnegativity constraints, and KKT complementary conditions.The D update separates positive and negative matrix parts after applying the KKT condition.
- Optimization procedure: Algorithm 1 initializes U, V, T, D, p, and q, precomputes the Laplacian matrix L, and repeatedly updates parameters until convergence.The updates cover D, U, V, T, and p,q.
- Convergence and prediction: Convergence is guaranteed because the nonnegative objective decreases monotonically during each iteration until reaching an optimal point.The algorithm then predicts labels for unlabeled news.
- Complexity: The overall time complexity sums initialization and fine-tuning costs, including O(nd +nld2 +rd +rm +n2) for D and O(tnd) for V per iteration.The reported costs also include updates for U, T, p, and q.
5 EXPERIMENTS
Experiments on BuzzFeed and PolitiFact evaluate TriFN against content-, social-context-, and combined-feature baselines, including component ablations and early-detection settings. TriFN consistently outperforms combined baselines, while both publisher and user social-context components contribute complementary information.
- Datasets and evaluation: Experiments use FakeNewsNet data from BuzzFeed and PolitiFact, with news labels, content, and social-context information.Evaluation reports Accuracy, Precision, Recall, and F1 using repeated train/test splits.
- Performance comparison: Social-context features outperform news-content features, while combining both feature types improves detection over either alone.The results identify complementary information between news content and corresponding social context.
- Performance comparison: TriFN consistently outperforms RST+Castillo and LIWC+Castillo across all evaluation metrics on both datasets.TriFN models publisher-news relations and user-news interactions alongside news content.
- Component analysis: Removing user social-engagement modeling reduces F1 by 5.2% and Accuracy by 6.1% on BuzzFeed, and F1 by 7.6% and Accuracy by 10.6% on PolitiFact.The ablation results support the importance of user social engagements in TriFN.
- Component analysis: Removing both publisher-partisan and user-engagement components reduces performance further, indicating complementary user-news and publisher-news embeddings.The component analysis concludes that both components contribute to performance improvement.
- Early fake news detection: F1 exceeds 80% within 48 hours on both datasets in early fake news detection.The experiment varies the available social-context delay from 12 to 96 hours; performance generally improves with longer delays for social-context methods.
6 RELATED WORK
Fake-news detection research uses both news content and social context, but TriFN is presented as the first approach to learn news features through publisher, content, and social-engagement relationships.
- Fake-news detection methods generally use news contents and social contexts.
- Content-based approaches extract linguistic and visual features, including writing styles, sensational headlines, and potentially fake images.
- Social-context approaches use user, post, and network features, including user profiles, characteristics, and credibility.
- Propagation-based approaches cannot be directly compared when only user actions are available because text-inferred propagation signals become ineffective.
- TriFN is presented as the first method to classify fake news by learning news features through publisher, news-content, and social-engagement relationships.
7 CONCLUSION AND FUTURE WORK
The paper concludes that TriFN models publisher, news, and social-engagement relationships for fake-news detection, with good early-dissemination performance. Future work targets early detection, psychological modeling of fake-news intention, and identifying low-quality or malicious users.
- Conclusion: TriFN models publisher, news, and social-engagement relationships and extracts features from publisher and user engagements while capturing their interrelationship.
- Conclusion: TriFN demonstrates effectiveness on real-world fake-news datasets and the importance of tri-relationship modeling for prediction.
- Future Work: The authors identify early detection, psychological modeling of fake-news intention, and detection of low-quality or malicious users as future directions.
8 ACKOWLEDGMENTS
The work was supported by NSF grant #1614576 and ONR grant N00014-16-1-2257.
- The work received support from NSF grant #1614576.
- The work received support from ONR grant N00014-16-1-2257.
- The acknowledgment states that the support was provided by, or in part by, the listed grants.