Skip to content
GCAN: Graph-aware Co-Attention Networks for Explainable Fake News Detection on Social Media

GCAN: Graph-aware Co-Attention Networks for Explainable Fake News Detection on Social Media

Authors: Yi-Ju Lu, Cheng-Te Li
Affiliation: National Cheng Kung University, Taiwan
Year: 2020 — arXiv:2004.11648

TL;DR

GCAN tackles fake news detection on Twitter with only source tweets and retweet sequences—no user comments or network structure required. The model combines graph convolution networks to capture user interactions with a dual co-attention mechanism to highlight suspicious retweeters and words. It achieves 87.7% accuracy on Twitter15 and 90.8% on Twitter16, outperforming prior work by ~18% while providing interpretable explanations.

Contributions

  • Addresses a realistic detection scenario: short-text tweets, no user response comments, no explicit network topology
  • Proposes GCAN, combining CNN/RNN propagation representations with graph-aware user interactions via GCN
  • Develops a dual co-attention mechanism (source-propagation and source-interaction) that provides explainability
  • Demonstrates early detection capability: achieves 90% accuracy with only 10 retweets
  • Extensive experiments on Twitter15 and Twitter16 show 18–20% improvement over state-of-the-art (CSI, dEFEND, CRNN)

Method

GCAN consists of five components:

User Characteristics Extraction: Extracts 10 features from each retweeting user's profile: self-description length, followers, following count, account age, verification status, location settings, temporal features, and retweet path distance.

Source Tweet Encoding: GRU processes word embeddings of the source tweet to capture sequential word dependencies.

User Propagation Representation: Models the retweet sequence using: - GRU-based: captures temporal patterns in user characteristics across the retweet chain - CNN-based: learns local sequential correlations among consecutive retweeters

Graph-aware Propagation Representation: Constructs a fully-connected graph where nodes are retweeting users and edge weights are cosine similarities between user feature vectors. A two-layer GCN then learns graph-aware node representations.

Dual Co-Attention Mechanism: Two attention modules: - Source-propagation co-attention: learns which source tweet words are attended to by which users in the retweet sequence, enabling explainability on both words and user order - Source-interaction co-attention: learns the correlation between source tweet words and user interactions (graph-aware features)

The final prediction concatenates outputs from both co-attention paths plus the GRU propagation representation through a feedforward network.

Results

Main Results (Table 3): - Twitter15: 87.67% accuracy, 82.50 F1 (vs. best competitor dEFEND: 73.83%, 65.41 F1) — 18.7% improvement - Twitter16: 90.84% accuracy, 75.93 F1 (vs. best competitor dEFEND: 70.16%, 63.11 F1) — 19.9% improvement - GCAN-G (without GCN) still outperforms prior work by 14–3%, showing the value of both graph structure and co-attention

Early Detection: With only 10 retweeters observed, GCAN achieves ~90% accuracy on Twitter15, significantly faster than baselines.

Ablation: Each component contributes meaningfully; removing co-attention or propagation representations causes the largest performance drops.

Explainability

The co-attention weights reveal: - Word-level: Fake news tends to attract attention to dramatic/sensational words ("breaking", "strict"), while real news highlights confirmatory language ("confirmed", "irrelevant") - User-level: Suspicious users for fake news are often unverified, have short account histories, and shorter paths to the source - Temporal: Fake news characteristics appear evenly distributed across retweet order, suggesting early detection is feasible but requires analyzing the full sequence

The model provides reasonable explanations without requiring long-text articles or user comments.

Connections

Notes

Strengths: - Realistic scenario (short tweets, no comments, no explicit network structure) reflects how most tweets actually propagate - Explainability is a strong practical feature for deployment; attention weights align with linguistic intuitions about fake news - Early detection with few retweets is valuable for timely intervention - Comprehensive comparison against 8 baselines and ablation studies validate each component

Limitations: - Evaluation limited to two Twitter datasets; generalization to other platforms (Facebook, Reddit) unclear - Fully-connected graph assumption may not reflect true user interactions; privacy-preserving alternative to explicit retweet graphs but potentially wasteful for large audiences - Hyperparameter choices (40 users, 32-dim embeddings) appear dataset-specific; sensitivity analysis missing - No user-level ground truth; cannot directly validate whether flagged users are actually suspicious

Future directions: Event-specific feature removal for robustness, cross-platform generalization, and extension to other short-text classification tasks.