Source-linked AI summary
FANG: Leveraging Social Context for Fake News Detection Using Graph Representation
Van-Hoang Nguyen, Kazunari Sugiyama, Preslav Nakov, Min-Yen Kan
TL;DR
Social media disinformation creates a need for scalable fake news detection beyond labor-intensive and modality-limited fact-checking. FANG models social entities and interactions with graph learning, improving representation quality and fake news detection while remaining robust with limited training data and supporting source factuality prediction.
Problem
Manual and evidence-based fact-checking requires considerable human effort and is not easily applicable to claims about images or videos, motivating improved contextual fake news detection.
Method
FANG is an inductive graph learning framework that models major social actors and their interactions to capture social structure and engagement patterns.
Results
FANG significantly improves fake news detection, remains robust with limited training data, and its learned representations support source factuality prediction.
Takeaways & Limitations
Modeling interactions between users, articles, and media enhances representation quality for fake news detection and source factuality prediction.
Takeaways & Limitations
Upstream textual encoding and stance-detection errors can propagate into FANG because entity and interaction features are constructed beforehand.
Abstract
from arXiv · showhide
We propose Factual News Graph (FANG), a novel graphical social context representation and learning framework for fake news detection. Unlike previous contextual models that have targeted performance, our focus is on representation learning. Compared to transductive models, FANG is scalable in training as it does not have to maintain all nodes, and it is efficient at inference time, without the need to re-process the entire graph. Our experimental results show that FANG is better at capturing the social context into a high fidelity representation, compared to recent graphical and non-graphical models. In particular, FANG yields significant improvements for the task of fake news detection, and it is robust in the case of limited training data. We further demonstrate that the representations learned by FANG generalize to related tasks, such as predicting the factuality of reporting of a news medium.
1 INTRODUCTION
Fake news on social media can disrupt public behavior and rationality, while existing verification and contextual approaches leave representation quality and minimally supervised settings underemphasized. FANG addresses this gap with an inductive graph-learning framework designed to represent social actors, interactions, and engagement patterns.
- Social-media disinformation can disturb social behavior, public fairness, and rationality, especially during elections or pandemics.
- Manual and automated fact-checking can be costly to scale, and textual-evidence methods do not readily handle image or video claims.
- Fake and factual news produce distinctive engagement patterns, including rapid recirculation and later denial posts for fake news.
- Earlier contextual models represented selected entities and interactions but placed limited emphasis on representation quality and minimally supervised settings.
- FANG proposes a graph representation covering major social actors and their interactions, using an inductive framework to capture social structure and engagement patterns.The framework is presented as improving representation quality for contextual fake news detection.
- The learned representations reportedly improve fake news detection, remain robust with limited training data, generalize to source-factuality prediction, and support explainability through recurrent-aggregator attention.
2 RELATED WORK
Related work represents social context through Euclidean features, network structures, and graph-learning frameworks, but differs in which entities, interactions, temporal patterns, and representation-learning goals it captures.
- Contextual fake-news detection research is categorized by how it represents and learns the social context.
- Euclidean approaches: Euclidean approaches flatten social context into vectors or matrices, using models ranging from Random Forests and SVMs to probabilistic models and LSTMs.
- Non-Euclidean approaches: Non-Euclidean approaches model user or source networks to capture structural features of social entities.
- Graph-based approaches: CSI combines dimensionality reduction of user co-sharing with RNN-derived engagement features, while TriFN omits stance, temporal engagement, and source-citation modeling.
- Graph-based approaches: Recent GNN models learn local and global structure but optimize fake-news detection alone, limiting robustness with scarce data and generalization to downstream tasks.
- Graph neural networks: GCNs face memory, heterogeneity, generalization, and transductive limitations, whereas GraphSage supports neighborhood aggregation, unsupervised proximity learning, and inductive inference.
3 METHODOLOGY
FANG constructs a heterogeneous social-context graph from news articles, sources, users, and timestamped interactions, then learns representations for fake-news detection. Its methodology combines textual entity features, stance and sentiment signals, graph structure, engagement temporality, and three jointly optimized objectives.
- Graph construction: The graph represents interactions through entity relations, labels, and, for publication and stance, relative timestamps.Stance labels include neutral support, negative support, deny, and report.
- Graph construction: FANG models questionable news articles, sources, social users, and their interactions as a social-context graph.Interactions relate entities through publication, citation, following, and user–article engagement, with timestamps for time-sensitive relations.
- Feature extraction: FANG constructs textual representations for articles and sources using TF.IDF and semantic vectors, while user features incorporate profile, activity, preference, and network signals.Article representations use unsupervised textual features; source vectors combine website text features with semantic information.
- Feature extraction: A stance detector classifies user replies relative to article titles into reporting, support, denial, and sentiment-specific support categories.The authors also construct a 2,527-pair dataset spanning 31 news events for stance detection and use a sentiment classifier to distinguish neutral from negative support.
- Representation learning: FANG learns entity representations with GraphSage and encodes news engagement sequences using a time-sensitive attention-based Bi-LSTM.Each engagement combines the elapsed time since publication, a one-hot stance vector, and the GraphSage representation of the participating user.
- Joint learning objectives: The framework jointly optimizes unsupervised Proximity, self-supervised Stance, and supervised Fake News Detection losses.The total objective is the linear combination L_total = L_prox. + L_stance + L_news.
4 EXPERIMENTS
The experiments evaluate FANG on a Twitter-based dataset using source, user, article, and interaction context, against content-only, Euclidean contextual, and graph-learning baselines. Fake-news detection is assessed with AUC, including temporal modeling comparisons.
- 4.1 Data: The dataset combines articles with their sources, engaged users, tweets, profile descriptions, and following relationships.Article truth labels come from Snopes and Politifact.
- 4.2 Fake News Detection Results: FANG is benchmarked against content-only, Euclidean contextual, and graph-learning models for fake news detection.The content-only baseline uses an SVM over TF.IDF news-content features.
- 4.2 Fake News Detection Results: 0.1993 absolute AUC improvement is achieved by FANG over the context-unaware baseline, compared with 0.1153 for CSI(-t).The comparison evaluates context-aware models using AUC.
- 4.2 Fake News Detection Results: 0.0339 AUC improvement is obtained by time-sensitive FANG over FANG(-t), compared with 0.0233 for CSI over CSI(-t).The variants differ in whether engagement representations include time(e).
- 4.2 Fake News Detection Results: 0.0501 AUC separates FANG(-t) from CSI(-t), while GCN exceeds CSI(-t) by 0.0386.These results support the effectiveness of the social graph representation relative to the Euclidean contextual baseline.
5 DISCUSSION
Discussion experiments test FANG with limited data, temporal engagement patterns, and representation quality. FANG remains effective across training sizes, distinguishes fake and real news through attention over time, and produces representations useful for clustering and source factuality prediction, while retaining practical and data-related limitations.
- 5.1 Limited Training Data (RQ1): FANG consistently outperforms both baselines at 10%, 30%, 50%, 70%, and 90% training-data sizes.Among graph models, FANG drops 11.11% from AUC 0.7518 at 90% to 0.6683 at 10%, while GCN drops 16.22%.
- 5.1 Limited Training Data (RQ1): 6.39% is the largest relative AUC margin of FANG over FANG(-s) as training data decreases, compared with 1.42% at 90% data.FANG(-s) removes the stance loss, highlighting the value of this self-supervised objective under limited data.
- 5.2 Contrastive Engagement Temporality (RQ2): 68.08% of fake-news attention falls within the first 12 hours, versus 48.01% for real news.Fake-news attention then falls to 18.83%, 4.14%, and approximately 9.04%; real-news attention remains 21.53% after two weeks.
- 5.3 Representation Learning (RQ3): FANG’s OPTICS clusters achieve homogeneity 0.051 from news representations, compared with 0.0006 for GCN.PCA plots show moderate fake–real collocation for FANG and little collocation for GCN.
- 5.3 Representation Learning (RQ3): 0.8049 AUC is achieved for source factuality classification, compared with 0.5842 for the baseline.FANG’s source representations use contextual relationships to distinguish media beyond textual similarity.
- 5.4 Scalability and Generalization: FANG directly infers embeddings for unseen nodes with a learned feature aggregator, avoiding the transductive handling required by prior approaches.This supports inference without maintaining or iteratively optimizing an embedding for every unseen node.
- 5.6 Limitations: Upstream textual-encoding and stance-detection errors can propagate because entity and interaction features are constructed before FANG.The dataset can also become obsolete when publication-time hyperlinks and social-media traces are no longer retrievable.
6 CONCLUSION AND FUTURE WORK
The paper concludes that modeling social context improves fake-news detection and that FANG enhances representations across fake-news and source-factuality tasks. It also identifies limited-data efficiency, temporal attention, and future multi-task learning as important directions.
- Conclusion: FANG captures social interactions among users, articles, and media to improve fake news detection and source factuality prediction.The conclusion frames representation quality and social context modeling as central to both tasks.
- Conclusion and Future Work: FANG remains effective with limited training data and uses explainable attention to capture temporal differences between fake and real news.Future work includes analyzing social-user representations and jointly learning fake-news detection, source factuality prediction, and echo-chamber discovery.