"This is Fake! Shared it by Mistake": Assessing the Intent of Fake News Spreaders

Authors: Xinyi Zhou, Kai Shu, Vir V. Phoha, Huan Liu, Reza Zafarani

Venue: ACM Web Conference 2022, April 25–29, Lyon, France — DOI

TL;DR

This work is the first to systematically assess whether individuals spread fake news intentionally or unintentionally. The authors propose an influence graph that captures psychological (internal) and social (external) factors driving unintentional spreading, validate it on two datasets with manually-annotated spreader intent, and demonstrate that the assessed intent significantly improves fake news detection (AUC ~0.93) and can inform platform interventions.

Contributions

• First intent modeling. Introduces a quantitative framework for assessing the degree to which fake news spreaders are intentional versus unintentional, grounded in social science theory.
• Influence graph. Proposes a directed, weighted, attributed graph that captures internal influence (a user's own past posts) and external influence (other users' posts) on each post-sharing action.
• Annotated datasets. Extends MM-COVID and ReCOVery with ground-truth intent labels (119 posts with high-confidence manual annotations; larger datasets with algorithmic annotation), enabling reproducible evaluation.
• Detection improvements. Shows that spreader intent features and the influence graph significantly enhance fake news detection, achieving ~0.93 AUC with hand-crafted features and 0.83–0.91 AUC with heterogeneous graph neural networks.

Method

The authors model fake news spreading through an influence graph \(G = (V, E, W)\) where vertices are posts and edges encode influence flow. Each post \(p_i = (A_i, c_i, t_i, u_i)\) is a tuple of articles shared, content (stance), time, and user. An edge from \(p_i\) to \(p_j\) exists if \(t_i < t_j\) and \(A_i \neq A_j\) (for same-user edges, internal) or \(u_i \neq u_j\) (external). Edge weight is:

\[W_{ij} = S(\bar{A}_i, A_j) \cdot S(c_i, c_j) \cdot T(\Delta t_{ij})\]

where \(S\) measures textual and visual similarity (using Sentence-BERT for short posts, Longformer for news articles), and \(T\) is an exponential decay function capturing user forgetting over time. The affected degree \(f_j = \sum_{(p_i, p_j) \in E} W_{ij}\) quantifies total influence on post \(j\); higher affected degree indicates more unintentional spreading.

For ground truth, they manually annotate 300 posts with expert + student annotators (κ = 0.61 agreement), then separately identify bots (Botometer), trolls (Bot Sentinel), and correctors (users who post corrections). The algorithmic annotation (bots, trolls, correctors → intentional; others → unintentional) achieves AUC 0.88–0.89 against manual labels.

Results

Validation of intent assessment: - Mean normalized affected degree for intentional spreaders: 0.55 (MM-COVID), 0.61 (ReCOVery) - For unintentional spreaders: 0.58, 0.62 (both significantly different, \(p \ll 0.001\) on MM-COVID) - Bots, trolls, and correctors all exhibit lower affected degree than unintentional spreaders (Fig. 5) - Spearman correlation ρ ∈ [−0.32, −0.24] (bots) and [−0.58, −0.36] (trolls) with affected degree

Fake news detection: - Using affected degree as a feature with XGBoost: AUC 0.935 (MM-COVID), 0.925 (ReCOVery) — beats dEFEND (0.90) and performs well with limited early-stage data (AUC >0.89 with first 10 posts) - Using influence graph with HetGNN: AUC 0.83–0.91 depending on labeled training fraction; outperforms subgraph variants

Connections

• Related to [[shao-spread-low-credibility-bots]] via bot behavior in propagation
• Related to [[vosoughi-spread-true-false-news]] on cascade structure and true vs. false news spreads
• Applies Reza Zafarani's prior work on survey and taxonomy (A Survey of Fake News: Fundamental Theories, Detection Methods, and Opportunities) to a new problem: spreader intent
• Motivated by Shifting attention to accuracy can reduce misinformation online on nudging and correction effectiveness

Notes

Strengths: First systematic quantitative treatment of spreading intent; strong social science grounding; well-designed influence graph capturing both content and temporal dynamics; reproducible annotation protocol with algorithmic-manual validation; clear practical utility (improves detection, informs platform policy on suspensions vs. nudging).

Limitations: The authors acknowledge that while they assess degree of intentionality via affected degree, binary classification (intentional/unintentional) requires a threshold, which is somewhat arbitrary. Future work should develop richer non-posting signals (comments, likes, follows) for intent estimation. Only considers Twitter; generalization to other platforms unclear.