Fake News Propagation Detection using Graph Attention Networks (GAT)

A research-grade implementation for detecting and analyzing fake news propagation patterns on social media using the FakeNewsNet dataset.

🎯 Project Objectives

• Find Key Spreaders – Identify influential users using attention weights
• Map Full Spread – Build propagation trees for each news item
• Build Simple, Strong Model – Efficient GAT without heavy architecture
• Get Better Results – Combine user influence, post content, and source credibility
• Explain Decisions – Interpret attention weights and propagation paths

📊 Dataset

FakeNewsNet: https://github.com/KaiDMML/FakeNewsNet

Components:

• News content (source, headline, body text, images/videos)
• Social context (user profiles, content, followers, followees)
• User-user and user-post interaction graphs

🏗️ Project Structure

majorProject/
├── data/                          # Raw and processed data
│   ├── raw/                       # FakeNewsNet raw JSON files
│   ├── processed/                 # Cleaned and preprocessed data
│   └── graphs/                    # Constructed graph structures
├── src/                           # Source code
│   ├── data/                      # Data ingestion and preprocessing
│   ├── features/                  # Feature engineering
│   ├── models/                    # GAT model definitions
│   ├── training/                  # Training scripts
│   ├── evaluation/                # Evaluation and metrics
│   └── visualization/             # Attention maps and graphs
├── notebooks/                     # Jupyter notebooks for exploration
├── configs/                       # Configuration files
├── experiments/                   # Training logs and checkpoints
├── outputs/                       # Results, predictions, reports
└── tests/                         # Unit tests

🚀 Setup Instructions

Prerequisites

• Python 3.10+
• uv package manager

Installation

1. Install uv (if not already installed):

curl -LsSf https://astral.sh/uv/install.sh | sh

2. Create virtual environment and install dependencies:

# Initialize project with uv
uv venv

# Activate virtual environment
source .venv/bin/activate  # On macOS/Linux

# Install dependencies
uv pip install -e .

3. Download FakeNewsNet dataset:

python scripts/download_dataset.py

🔬 Research Pipeline

Stage 1: Data Ingestion and Preprocessing

python src/data/preprocess.py --input data/raw --output data/processed

Stage 2: Graph Construction and Feature Engineering

python src/features/build_graph.py --config configs/graph_config.yaml

Stage 3: Model Training

python src/training/train_gat.py --config configs/model_config.yaml

Stage 4: Evaluation and Analysis

python src/evaluation/evaluate.py --checkpoint experiments/best_model.pt
python src/visualization/generate_reports.py

📈 Expected Results

Metrics

• Accuracy
• F1-Score
• AUC-ROC
• Explanation Fidelity

Visualizations

• Propagation graphs
• Attention-based influence maps
• Confusion matrices
• Performance summaries

🧪 Usage Examples

Training a GAT Model

from src.models.gat_model import FakeNewsGAT
from src.training.trainer import GATTrainer

model = FakeNewsGAT(
    in_channels=768,
    hidden_channels=128,
    num_classes=2,
    num_layers=3
)

trainer = GATTrainer(model, train_loader, val_loader)
trainer.train(epochs=100)

Analyzing Key Spreaders

from src.evaluation.explainability import AttentionAnalyzer

analyzer = AttentionAnalyzer(model, graph_data)
key_spreaders = analyzer.identify_influential_users(top_k=20)
analyzer.visualize_propagation_tree(news_id='article_123')

📚 Tech Stack

• PyTorch Geometric: Graph neural networks
• Transformers (HuggingFace): BERT embeddings
• NetworkX: Graph construction and analysis
• scikit-learn: Classical ML metrics and preprocessing
• matplotlib/seaborn: Visualization

🔍 Interpretability

All predictions are accompanied by:

• Attention weight distributions
• Top influential users in propagation chains
• Propagation path visualizations
• Feature importance analysis

📝 Notes

• Purpose: Research and analysis only
• Focus: Code efficiency and model accuracy
• Interpretability: Attention weights logged and visualized for all predictions

📄 License

MIT License - For research purposes only

🤝 Contributing

This is a research project. For questions or collaboration, please open an issue.

📧 Contact

[Your contact information]