FANG: Leveraging Social Context for Fake News Detection Using Graph Representation

Van-Hoang Nguyen; Kazunari Sugiyama; Preslav Nakov; Min-Yen Kan

doi:10.1145/3340531.3412046

In brief

FANG proposes a graph neural network framework that represents fake news detection as a heterogeneous network of users, news articles, and sources with multiple interaction types (stance, publication, citation, followership). Using GraphSage with attention-based temporal aggregation, FANG learns high-quality representations that capture engagement patterns and temporal dynamics distinctive between fake and real news, achieving 0.75 AUC and showing robustness under limited training data.

Contributions

Heterogeneous graph representation of social context modeling all major entities (users, news, sources) and interactions (stance, publication, citation, followership) including temporal dynamics.
FANG learning framework: An inductive graph neural network combining GraphSage for structural learning with Bi-LSTM+attention for temporal engagement aggregation.
Multi-objective learning: Joint optimization of proximity loss (for echo chambers and polarization), stance loss (self-supervised), and fake news detection loss.
Representation quality focus: Demonstrates that FANG learns generalizable representations verified on downstream source factuality prediction (0.80 AUC) and intrinsic clustering evaluation.
Explainability via attention: Attention mechanism over temporal windows reveals that the model appropriately weights early engagements for fake news (68% of attention in first 12 hours) vs. sustained attention for real news.

Method

Graph construction encodes news articles using TF-IDF + GloVe embeddings; sources from Homepage and About Us text; users from profile descriptions. Stance detection between user posts and article titles uses a fine-tuned RoBERTa classifier (0.886 accuracy) on a novel 2,527-sample annotated dataset, differentiating between neutral support, negative support, deny, and report stances.

FANG learns three representations:

Structural via GraphSage sampling and aggregating node neighborhoods over the heterogeneous graph
Temporal via Bi-LSTM encoding user engagement sequences with timestamps and stance labels, weighted by attention that measures engagement relevance to the article content
Combined as the sum of temporal (Bi-LSTM output) and structural (GraphSage output) representations for news nodes

The model optimizes a sum of three losses: (i) Proximity loss enforcing nearby entities in the graph have similar embeddings (motivated by echo chambers and polarization), (ii) Stance loss pulling user and article embeddings close in a multi-head stance projection space when the user expresses that stance, and (iii) Fake News Loss (supervised cross-entropy) predicting article veracity from the concatenation of the article and source representations.

Results

On a Twitter dataset of 448 fake and 606 real articles with 54k users and 442 sources:

Fake news detection: FANG achieves 0.7518 AUC, outperforming CSI (Euclidean contextual model, 0.6911) and GCN (graph-only, 0.7064). Temporal modeling adds 0.0339 AUC over FANG without time.
Limited data robustness: FANG drops 11.11% AUC from 90% to 10% training data (0.7518 → 0.6683); GCN drops 16.22%; CSI drops 7.93%. Ablating stance loss increases this to 6.39% relative margin at 30% data.
Source factuality prediction: FANG representations achieve 0.8049 AUC predicting source credibility (high/low/mixed), vs. 0.5842 for text-only baseline, demonstrating representation generalization.
Intrinsic clustering: FANG representations yield 0.051 homogeneity on OPTICS clustering by factuality label vs. 0.0006 for GCN, showing tighter collocation within fake and real classes.

Notes

Strengths:

First work to comprehensively model all major social actors (users, sources, articles) with their heterogeneous interactions in a single GNN framework.
Novel stance-annotated dataset (2,527 samples, Cohen's κ = 0.78) is a valuable resource.
Attention mechanism provides genuine interpretability: the 68% attention to first-12-hour engagements for fake news aligns with known patterns of rapid viral spread of sensational false content.
Robustness to limited training data suggests the learned social context representations capture fundamental signal.
Inductive approach scales better than transductive methods; inference complexity depends only on neighborhood size, not full graph.

Limitations and future work:

Errors in upstream tasks (text encoding, stance detection) propagate; end-to-end joint optimization suggested.
Dataset can become obsolete as hyperlinks and social traces disappear; affects reproducibility and real-world deployment.
Focuses on social context; content features are represented only via TF-IDF/GloVe; multimodal false claims may be underrepresented.
No comparison to more recent GNN variants (e.g., GAT, heterogeneous graph frameworks).

Fit for GUIDE: This is a canonical method paper that bridges representation learning and fake news detection. The heterogeneous graph formulation, temporal aggregation, and multi-loss training approach are widely cited and have influenced subsequent work in the field. Strongly in scope.