Propagation-based fake news detection

Detection methods that analyze how news spreads through social networks. Rather than examining content or source credibility, propagation-based approaches examine the structure and dynamics of how claims diffuse: who shares them, how quickly, how far, and the branching patterns of retweets and shares.

Key insights

• False news spreads differently: Falsehood typically diffuses farther, faster, deeper, and more broadly than truth The Spread of True and False News Online
• Cascade structure as signal: Tree-like retweet chains contain information about claim veracity; self-reinforcing cascades, user diversity, and depth patterns distinguish true from false claims
• Temporal dynamics matter: Early retweet velocity, time to first shares, and acceleration curves vary between true and false news
• Network topology features: Structural properties like maximum breadth, retweet depth, and user rank distributions are predictive

Complementary to other approaches

Propagation-based detection works alongside: - Content-based detection (linguistic and visual features) - Credibility assessment (source and author reputation) - Feature engineering (representing cascade structure as machine-learning features)

Key papers in this wiki

• Shu, Bernard & Liu (2018) — Studying Fake News via Network Analysis — comprehensive survey of network types (homogeneous: friendship, diffusion, credibility; heterogeneous: knowledge, stance, interaction) and detection methods including temporal RNN models for engagement sequences, credibility network propagation via belief propagation, and knowledge network matching for fact-checking
• Lu & Li (2020) — GCAN: Graph-aware Co-Attention Networks for Explainable Fake News Detection: Detects fake news from short tweets using only retweet user sequences and profiles; models retweet propagation with CNN/GRU-based representations and constructs fully-connected graphs where edges weight cosine similarity between user features; applies GCN and dual co-attention to identify suspicious propagation patterns and highlight informative words; achieves 87.7% and 90.8% accuracy on Twitter15/16 with strong early detection (90% accuracy with 10 retweets).
• Monti et al. (2019) — Fake News Detection on Social Media using Geometric Deep Learning: applies graph convolutional networks to cascade structures, showing that propagation-based features (network topology, cascade shape, temporal dynamics) achieve 92.7% ROC AUC and enable early detection within 1–2 hours; demonstrates language independence of propagation patterns and their robustness to adversarial attacks compared to content-based methods
• Sharma et al. (2018) — Combating Fake News: A Survey on Identification and Mitigation Techniques: comprehensive survey of propagation-based detection methods including propagation tree kernels (Ma et al. 2017), propagation tree neural networks (Ma et al. 2018), SEIZ process modeling (Jin et al. 2013), temporal variation features, and response analysis; discusses computational trade-offs (tree kernel methods are expensive at scale) and early detection challenges
• Ma et al. (2017) — Detect Rumors in Microblog Posts Using Propagation Structure via Kernel Learning: kernel-based approach (PTK/cPTK) to detect rumors by measuring structural similarity between propagation trees; soft-matches subtrees to capture high-order patterns; demonstrates importance of structural signals beyond temporal features; achieves 75% accuracy on Twitter15 and superior early detection at 24-hour mark.
• Castillo et al. (2011) — Information Credibility on Twitter: early work showing retweet tree structure (depth, breadth, maximum level size) predicts content credibility; demonstrates that propagation patterns alone achieve 75.8% F1 for credibility classification, and that credible information originates from fewer, established users while false claims come from diverse but low-reputation sources.
• Ma et al. (2018) — Rumor Detection on Twitter with Tree-structured Recursive Neural Networks: applies recursive neural networks to propagation trees, learning joint content-structure representations; demonstrates that local patterns in reply structure (supportive replies to denials, questioning replies to affirmations) carry veracity signals; achieves 72.3% accuracy with 4× faster early detection.
• Bian et al. (2020) — Bi-Directional Graph Convolutional Networks for Rumor Detection: First GCN-based rumor detection combining top-down propagation patterns and bottom-up dispersion patterns; source-post feature enhancement amplifies early information; achieves 96.1% Weibo, 88.6% Twitter15, 88.0% Twitter16; demonstrates GCN scalability over tree-based methods for large cascades.
• Shu et al. (2019) — Hierarchical Propagation Networks for Fake News Detection: constructs hierarchical propagation networks at both macro-level (retweet cascades) and micro-level (reply conversations); extracts structural, temporal, and linguistic features showing fake news spreads deeper, contains more bot-amplified cascades, has shorter lifespan, and generates more negative sentiment; hierarchical network features (HPNF) achieve F1 > 0.80 on FakeNewsNet, outperforming macro-level-only approaches (STFN) by demonstrating complementarity of micro-level signals.
• Zhou et al. (2019) — WSDM Tutorial on Fake News Detection: presents propagation-based detection as one of four unified perspectives, encompassing spreader networks, retweet cascades, and temporal dynamics.
• The Spread of True and False News Online (2017) — foundational large-scale empirical study of diffusion patterns
• A Survey of Fake News (2020) — systematic treatment of propagation-based detection as one of four perspectives
• Network-based Fake News Detection: A Pattern-driven Approach (2019)
• Monti et al. (2019) — Fake News Detection on Social Media using Geometric Deep Learning: applies graph convolutional networks to cascade structures, showing that propagation-based features (network topology, cascade shape, temporal dynamics) achieve 92.7% ROC AUC and enable early detection within 1–2 hours; demonstrates language independence of propagation patterns and their robustness to adversarial attacks compared to content-based methods

Open questions

• How do algorithmic feeds and recommendation systems alter traditional propagation patterns?
• What is the relative importance of early-stage cascade features vs. long-term patterns for early detection?
• How do propagation patterns differ across platforms (Twitter, Facebook, TikTok)?
• Can intervention (labels, friction, prebunking) measurably change cascade shapes before they go viral?