Research map / Graph learning / Graph self-supervised learning
Propagation Tree Is Not Deep: Adaptive Graph Contrastive Learning Approach for Rumor Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that rumor propagation trees are mostly wide and shallow, and proposes RAGCL, a graph contrastive learning method with centrality-guided adaptive node dropping, attribute masking and edge dropping for rumor detection.
Problem
Graph-based rumor detection models assume rumor propagation trees are deep and learn sequential stance features along branches, but real RPTs are wide with most replies being shallow 1-level replies.
Approach
Adaptive graph augmentation of RPTs guided by node centrality (degree, betweenness, PageRank), with three principles (exempt root, retain deep reply nodes, preserve lower-level nodes in deep sections), feeding a GCN-based unsupervised contrastive regularizer added to a supervised rumor classification loss.
Key results
- RAGCL reaches 0.962 accuracy on Weibo and 0.894 on DRWeibo, above the best baseline DDGCN (0.948 and 0.878).
- On Twitter15 RAGCL achieves 0.867 accuracy versus 0.846 for GACL; on Twitter16 it achieves 0.905 versus 0.893 for DDGCN.
- Statistics show 1-level replies make up 64-78% of replies across the four labeled datasets, and only 6.4-10.8% of 1-level replies receive further replies.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RAGCL instead of DDGCN | rumor detection · Weibo | accuracy: 0.948 ±0.004 -> 0.962 ±0.005 | Helped |
| use RAGCL instead of DDGCN | rumor detection · DRWeibo | accuracy: 0.878 ±0.005 -> 0.894 ±0.004 | Helped |
| use RAGCL instead of GACL | rumor detection · Twitter15 | accuracy: 0.846 ±0.007 -> 0.867 ±0.005 | Helped |
| use RAGCL instead of DDGCN | rumor detection · Twitter16 | accuracy: 0.893 ±0.004 -> 0.905 ±0.003 | Helped |
| add node dropping and edge dropping instead of supervised GCN alone (ablation) | RAGCL · Weibo | accuracy: 0.927 -> 0.962 | Helped |
| add node dropping and edge dropping instead of supervised GCN alone (ablation) | RAGCL · DRWeibo | accuracy: 0.844 -> 0.894 | Helped |
| add node dropping and edge dropping instead of supervised GCN alone (ablation) | RAGCL · Twitter15 | accuracy: 0.822 -> 0.864 | Helped |
| add node dropping and edge dropping instead of supervised GCN alone (ablation) | RAGCL · Twitter16 | accuracy: 0.846 -> 0.902 | Helped |
| replace adaptive node dropping and attribute masking with random augmentation (ablation) | RAGCL · Weibo | accuracy: 0.940 -> 0.953 | Hurt |
| replace adaptive node dropping and attribute masking with random augmentation (ablation) | RAGCL · DRWeibo | accuracy: 0.861 -> 0.892 | Hurt |
| use attribute masking in Chinese datasets (ablation) | RAGCL · Weibo and DRWeibo | accuracy | Hurt |
| use degree rather than PageRank centrality (ablation) | RAGCL · Weibo | accuracy: 0.962 -> 0.953 | Hurt |
Surprising findings
- Eliminating a growing fraction of unresponded 1-level replies leaves performance steady or slightly improved, suggesting these replies may be noise.
- Attribute masking reduces performance on the Chinese datasets (Weibo and DRWeibo).
Limitations
- Gains over baselines on Twitter datasets are modest, and the paper reports no significance tests beyond standard deviations.
- Best reported RAGCL results use the best centrality and augmentation combination per dataset, so the setting is selected on test results.
- The supplementary material with hyperparameter and setting details is not included in the provided text.
Experimental setup
Datasets: Weibo DRWeibo Twitter15 Twitter16 UWeibo UTwitter
Benchmarks: Weibo DRWeibo Twitter15 Twitter16
Metrics: accuracy precision recall F1
Baselines: PLAN BiGCN UDGCN GACL DDGCN
Builds on: BiGCN graph contrastive learning with graph augmentation GACL
Closest papers by meaning
- Backdoor Attack on Propagation-based Rumor Detectors (AAAI 2025 · similarity 0.86)
- GIN-SD: Source Detection in Graphs with Incomplete Nodes via Positional Encoding and Attentive Fusion (AAAI 2024 · similarity 0.85)
- Enhancing Robustness of Graph Neural Networks on Social Media with Explainable Inverse Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- Hierarchical and Incremental Structural Entropy Minimization for Unsupervised Social Event Detection (AAAI 2024 · similarity 0.84)
- GAMC: An Unsupervised Method for Fake News Detection Using Graph Autoencoder with Masking (AAAI 2024 · similarity 0.83)
- Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.