Label propagation¶
Label propagation is a semi-supervised learning technique that leverages graph structure to infer labels for unlabeled data by propagating information from labeled nodes through connections. In the context of fake news detection, label propagation can spread veracity signals across related news items or user networks, improving detection when labeled training data is scarce.
Key observations¶
Graph structure matters: Label propagation assumes that connected nodes (related news items or users) should have similar labels. This assumption holds well for misinformation detection when the graph captures semantic similarity or propagation networks.
Weak pseudo-labels can mislead: Direct propagation of weak pseudo-labels can amplify errors. Masking mechanisms or confidence-based filtering are necessary to prevent label leakage and improve robustness.
Multimodal integration enhances propagation: When combined with multimodal representations that capture both textual and visual consistency, label propagation can leverage cross-modal signals to improve label quality and propagation effectiveness.
Related topics¶
- Graph Neural Networks — neural network architectures for graph-structured data
- Semi Supervised Learning — learning from partially labeled data
- Weak supervision — learning from noisy or incomplete labels
- Pseudo-labeling — generating labels for unlabeled data
- Information diffusion in social networks — propagation of information through networks
Key papers¶
- Hu, Hu & Zhang (2025) — Synergizing LLMs with Global Label Propagation for Multimodal Fake News Detection — integrates LLM-based pseudo-labeling with masked label propagation for improved detection
- Zhu & Ghahramani (2002) — Learning from Labeled and Unlabeled Data with Label Propagation — foundational work introducing label propagation for semi-supervised learning