Prominent Features of Rumor Propagation in Online Social Media

Contributions

• Novel temporal model (PES): Extends the SpikeM epidemiological model by introducing periodic external shocks, explaining why rumors generate repeating bursts of activity while non-rumors decay monotonically.
• Multimodal feature analysis: Comprehensively examines 91 features across three dimensions (11 temporal, 15 structural, 65 linguistic), grounded in social psychology theory.
• Foundational rumor dataset: Annotated 104 viral events (47 rumors, 55 non-rumors) from 3.5 years of near-complete Twitter data with high inter-coder agreement (ICC=0.992); publicly released for community use.
• Superior classification performance: Random forest classifier with proposed features achieves F1=0.893 (accuracy 90%, precision 93.5%, recall 89.2%), significantly outperforming baseline credibility features (F1=0.788).
• Temporal-feature primacy: Demonstrates that periodicity of external shock (qp) is the single most predictive feature, suggesting rumors succeed by generating repeated stimulus rather than sustained information flow.

Method

Data

• Source: Twitter near-complete archive covering 3.5 years (March 2006–August 2009): 54M user profiles, 1.9B follow links, 1.7B public tweets
• Topic collection: Investigated rumor archives (snopes.com, urbanlegends.about.com, pcmag.com) and news websites (times.com) to identify 130 topics (70 rumors, 60 non-rumors) occurring within dataset timeframe
• Annotation: Four coders classified each topic; selected topics with ≥3 coders' majority agreement, yielding 68 rumors + 57 non-rumors
• Filtering: Retained 102 topics with ≥60 tweets; final dataset: 47 rumors, 55 non-rumors with ICC=0.992

Temporal features (11 variables from PES model)

Proposed Periodic External Shocks (PES) model: Extends SpikeM by decomposing shock into baseline and periodic components. For each event, fitted model parameters capture:

• N: total population of available users
• β: infection probability (retweet propensity)
• nb: starting time of breaking news
• Sb: initial shock strength
• ε: background noise
• pa, ps: interaction periodicity (amplitude, phase) modeling daily cycles
• qa, qp, qs: external shock periodicity (amplitude, period, phase) capturing rumor-specific cycles

Result: rumors show qp < non-rumors (shorter periodic cycle), explaining distinctive multiple-spike pattern.

Structural features (15 variables)

Extracted three network types:

• Friendship network: induced subgraph of follower-followee relationships among event participants
• LCC (largest connected component): core network structure
• Diffusion network: inferred from temporal posting order (user A→B if B follows A and posts after A about the topic)

Features: degree distributions, clustering coefficient, density, fraction of isolated spreaders, fraction of low→high degree diffusion (less-influential→more-influential flow).

Linguistic features (65 variables)

Used LIWC (Linguistic Inquiry and Word Count) sentiment analyzer on consolidated tweet sets. Tested major categories (social, affective, cognitive, perceptual, biological) and subcategories (e.g., positive emotion, negation, certainty, tentative, anxiety, insight).

Results

Feature selection & importance

Random forest + logistic regression identified 11 significant features (t-test, p<0.05):

Temporal (3 features):

• qp (external shock period) — N ✓✓ (highest importance, non-rumors have longer cycles, i.e., single spike)
• qs (shock periodicity offset) — N ✓✓
• ps (interaction periodicity offset) — N ✓

Structural (5 features):

• Cg (clustering, friendship network) — R (rumor spreaders more clustered)
• Dl (density, LCC) — R
• Cl (clustering, LCC) — R
• Sd (fraction of isolated nodes) — R ✓ (rumors have more singletons; followers neglect)
• Fd (low→high degree diffusion) — R ✓ (rumor diffusion flows from less to more influential users)

Linguistic (3 features):

• posemo (positive emotion words: love, nice, sweet) — N ✓ (non-rumors have more)
• negate (negation: no, not, never) — R ✓ (rumors have more; listeners express skepticism)
• social (social words: mate, talk, they, child) — R (rumor-related discussion is more social)

Classification results

Compared three feature sets using decision tree, random forest, and SVM with 2-fold cross-validation (1000 iterations):

Key insight: S1 alone outperforms combined set C (F1 0.893 vs 0.878), indicating proposed features are not redundant with baseline and actually capture complementary signals; adding baseline features introduces noise.

Interpretation

Temporal dominance: Periodicity of external shock is the single strongest feature. Unlike baseline features, which measure static attributes (user follower count, URL presence), the PES model captures the dynamics of rumor cycles. Rumor audience may neglect messages without credibility evidence, requiring new "shocks" (exogenous events, celebrity retweets) to reignite spreading. True information spreads via steady institutional channels (news orgs) and thus shows single-peak decay.

Structural patterns: Rumors concentrate spreading among loosely-connected communities (high Sd—singletons). The diffusion flow from low→high degree nodes (Fd) reflects the surprising finding that rumors reach influential users through less-influential followers, whereas true information typically originates from high-credibility sources.

Linguistic signals: Rumors trigger more skepticism (negation, cognitive words) and less positive sentiment than true news—users interrogate rumorous claims. Social-process words indicate rumor discussion is interpersonal (gossip) rather than broadcast.

Notes

Strengths:

• Theoretical grounding: Features derive from social psychology (Allport, Rosnow, Shibutani, Sunstein on rumor dynamics) rather than ad-hoc engineering
• Large-scale validation: Near-complete Twitter archive over 3.5 years; 102 events with careful annotation (ICC=0.992)
• Novel temporal model: PES elegantly captures the key distinguishing feature (periodic spikes) and outperforms simpler alternatives
• Dataset contribution: Publicly released annotated rumors enabling reproduction and follow-up work
• Practical significance: Temporal features are often more accessible than structural features (don't require full network history), lowering real-time detection barriers
• Comprehensive feature evaluation: Systematic comparison across three domains (temporal, structural, linguistic) with statistical significance testing

Limitations:

• Language and platform: Twitter-centric study; generalization to Weibo, Facebook, Reddit unclear; only English
• Historical data: 2006–2009 archive may not reflect modern platform dynamics (verification badges, reduced RT friction, algorithmic feeds)
• Ground truth: Rumor annotation depends on archival websites (snopes.com, urbanlegends.about.com); some borderline cases or contemporary rumors not covered
• Feature engineering: LIWC-based linguistic features are lexicon-dependent; modern semantic embeddings might capture deeper signals
• Binary classification: Does not model finer veracity gradations (confirmed true, unconfirmed, false, denied)—extends to only true/false/unknown in post-hoc judgment
• Causal claims: Observational features; cannot claim that periodic shocks cause rumors to spread or vice versa

Historical context:

This paper established temporal dynamics (particularly periodic bursts) as a core signal for rumor detection, influencing the field's subsequent focus on time-series and cascade-based methods. Prior work emphasized static credibility cues (user reputation, URL presence); this work showed that when information appears relative to time and events matters as much as who spreads it. Later work extended this to RNNs (Ma et al. 2016) and GNNs (Bian et al. 2020) while retaining the core insight that rumor cycles differ from true-information cascades.