Social-context-based fake news detection

Social-context approaches to fake news detection exploit signals from the social environment in which news is shared — user profiles, engagement patterns, propagation structure, and network topology — rather than (or in addition to) the news content itself. The motivation is that fake news, crafted to mimic real news textually, may be harder to distinguish by content alone but leaves traces in how it spreads and who shares it.

Social-context models are broadly divided into: - Stance-based: aggregate users' expressed opinions or stances toward a news item to infer veracity. - Propagation-based: model the diffusion network or temporal pattern of shares/retweets. - Profile-based: characterize the demographics and behaviors of users who engage with news.

Key papers and datasets

Dataset paper: - Shu et al. (2018) — FakeNewsNet: Comprehensive multi-dimensional repository integrating news content, social context (user engagement, network structure), and spatiotemporal information; contains PolitiFact (12,911 articles) and GossipCop (22,140 articles) with rich social signals enabling downstream social-context detection research.

Detection approaches: - Shu et al. (2019) — Beyond News Contents: The Role of Social Context for Fake News Detection (TriFN): tri-relationship embedding approach modeling publisher-news and user-news relationships simultaneously; incorporates publisher partisan bias and user credibility; achieves 4–6% F1 improvement over content+profile baselines on FakeNewsNet; demonstrates early detection (>80% F1 within 48 hours). - Ma et al. (2018) — Rumor Detection on Twitter with Tree-structured Recursive Neural Networks: propagation-based approach using recursive neural networks on thread structures; jointly models content and structure to detect rumors; achieves 72.3% accuracy on Twitter15. - Zhou & Zafarani (2019) — Network-based Fake News Detection: A Pattern-driven Approach: four network-structural patterns (More-Spreader, Farther-Distance, Stronger-Engagement, Denser-Network) operationalized as 138 features; RF achieves 0.929/0.932 accuracy/F₁ on PolitiFact, outperforming content-based and hybrid baselines without reading any content. - Shu et al. (2019) — Hierarchical Propagation Networks for Fake News Detection: propagation-based approach extracting structural, temporal, and linguistic features from both macro-level retweet cascades and micro-level reply conversations; shows fake news exhibits deeper cascades, more bot amplification, shorter lifespan, and more negative sentiment; hierarchical features achieve F1 > 0.80, outperforming prior single-level propagation approaches. - Shu et al. (2019) — The Role of User Profiles for Fake News Detection: profile-based approach; UPF beats content baselines on FakeNewsNet. - Shu et al. (2019) — dEFEND: Explainable Fake News Detection: jointly encodes news content and user comments via hierarchical attention and sentence-comment co-attention; outperforms content-only and comment-only baselines on FakeNewsNet; provides explainability via top-k sentences and comments. - Yang et al. (2020) — CHECKED: provides one of the richest available propagation graphs — 1,868,175 Weibo reposts and 1,185,702 comments with full thread content, hashed user IDs, and timestamps — over Chinese COVID-19 microblogs; propagation-based methods are not evaluated in the paper, leaving a direct extension opportunity. - Monti et al. (2019) — Fake News Detection on Social Media using Geometric Deep Learning: graph convolutional network approach integrating user profiles (geolocation, verification status, age), user activity (favorites, retweets, lists), network structure (social connections), and propagation patterns (cascade trees, retweet timestamps); achieves 92.7% ROC AUC on Twitter; shows user profile and network/spreading features are most important (90% AUC when combined) while content features contribute marginally; enables early detection within 1–2 hours.

Connections

• User profiles are one primary source of social-context features.
• Political bias emerges as an important signal within social-context analysis.
• Feature engineering covers the broader methodology for constructing social-context feature vectors.
• Contrasts with content-based approaches such as RST discourse parsing or LIWC psycholinguistic features.