Graph Based Detection

Detection methods that model information propagation and social networks as graphs, then apply graph neural networks (GNNs), spectral analysis, or topological feature extraction to distinguish fake from real news. These approaches leverage the insight that fake and real news exhibit different cascade structures: depth, breadth, balance, and connectivity patterns.

Key insights

• Network topology is diagnostic: Propagation trees, retweet graphs, and social networks exhibit structural patterns that correlate with veracity; fake news tends to spread in deeper, more unbalanced structures.
• Graph neural networks learn structure: GNNs such as graph convolutional networks (GCNs) and graph attention networks automatically learn feature representations from graph structure without hand-engineered features.
• Spectral methods provide interpretability: Spectral bounds and eigenvalue analysis translate abstract graph properties into interpretable cascade characteristics (branching capacity, cohesion, diffusion dynamics).

Methods and approaches

• Graph neural networks: GCNs, RNNs on graph sequences, graph attention networks applied to cascade graphs for end-to-end learning.
• Spectral analysis: Eigenvalue-based bounds and spectral graph theory to derive interpretable structural properties and classification features.
• Kernel methods: Graph kernels (e.g., propagation tree kernels) measuring structural similarity between cascades.
• Topological features: Hand-engineered features from cascade depth, breadth, maximum degree, and other graph-theoretic properties.

Key papers

• Cai & Zafarani (2026) — Spectral Analysis of Fake News Propagation: Derives 35 spectral bounds capturing branching capacity, cascade scale, structural cohesion, propagation span, and diffusion dynamics; shows spectral representations achieve competitive classification performance while providing interpretability.
• Monti et al. (2019) — Fake News Detection on Social Media using Geometric Deep Learning: Applies GCNs to cascade structures, achieving 92.7% ROC AUC and demonstrating language independence of propagation-based detection.
• Bian et al. (2020) — Bi-Directional Graph Convolutional Networks for Rumor Detection: First GCN combining top-down propagation and bottom-up dispersion patterns; achieves 96.1% on Weibo.

See also

• Propagation-based fake news detection for broader treatment of cascade-structure-based approaches
• Spectral Analysis for graph spectral theory and eigenvalue-based methods
• Network Based Misinformation Detection for multi-modal network analysis