Research map / Graph learning / Graph self-supervised learning
GAMC: An Unsupervised Method for Fake News Detection Using Graph Autoencoder with Masking
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.
Proposes GAMC, an unsupervised graph autoencoder with node-feature masking, edge dropping, and a reconstruction plus contrastive loss for fake news detection on FakeNewsNet propagation graphs.
Problem
Fake news detection relies on large labeled datasets for graph-based methods that model social propagation context, and labeling is costly.
Approach
Applies a masked graph autoencoder (GIN encoder, GIN decoder with re-masking) to news propagation graphs, trained on two augmented views with a reconstruction MSE loss and a cosine-similarity contrast term, then classifies pooled graph embeddings with an SVM.
Key results
- On PolitiFact, GAMC reaches 0.828 accuracy and 0.823 F1, above the best unsupervised baseline (UMD)^2 at 0.802 accuracy.
- On GossipCop, GAMC reaches 0.941 accuracy and 0.937 F1, versus 0.792 accuracy for the best unsupervised baseline (UMD)^2.
- GAMC is competitive with supervised methods, trailing BiGCN (0.951 acc) and GACL (0.867 acc on PolitiFact) on some metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GAMC instead of (UMD)^2 | Unsupervised fake news detection · PolitiFact | accuracy: 0.802 -> 0.828 | Helped |
| Use GAMC instead of (UMD)^2 | Unsupervised fake news detection · GossipCop | accuracy: 0.792 -> 0.941 | Helped |
| Use GAMC instead of GACL | Fake news detection · PolitiFact | accuracy: 0.867 -> 0.828 | Hurt |
| Use GAMC instead of BiGCN | Fake news detection · GossipCop | accuracy: 0.951 -> 0.941 | Hurt |
| Add node feature masking and edge dropping (ablation) | GAMC · PolitiFact | accuracy: 0.793 -> 0.828 | Helped |
| Add node feature masking and edge dropping (ablation) | GAMC · GossipCop | accuracy: 0.905 -> 0.941 | Helped |
| Add reconstruction loss (ablation) | GAMC · PolitiFact | accuracy: 0.763 -> 0.828 | Helped |
| Add reconstruction loss (ablation) | GAMC · GossipCop | accuracy: 0.889 -> 0.941 | Helped |
| Add contrastive loss and a second augmented graph (ablation) | GAMC · PolitiFact | accuracy: 0.808 -> 0.828 | Helped |
| Add contrastive loss and a second augmented graph (ablation) | GAMC · GossipCop | accuracy: 0.924 -> 0.941 | Helped |
| Set mask rate to 0.5 and edge drop rate to 0.2 (ablation) | GAMC · PolitiFact and GossipCop | performance | Helped |
Surprising findings
- The paper reports that removing the reconstruction loss lowers PolitiFact accuracy to 0.763, and that the variant without the contrast loss (GAMC-L_con) scores 0.808 on PolitiFact, which is higher than the full model on GossipCop only in the reverse direction; the contrast term is therefore not unif
- The contrastive loss as written is a cosine similarity that is minimized with a negative sign in the total loss, which is unusual for contrastive learning and is not clearly justified.
Limitations
- Evaluated only on two FakeNewsNet subsets (PolitiFact and GossipCop) with text-only node features.
- Still uses a supervised SVM classifier on learned embeddings, so labels are needed for the final classification step.
- Does not outperform the strongest supervised graph baselines such as BiGCN on GossipCop and GACL on PolitiFact.
Experimental setup
Datasets: FakeNewsNet PolitiFact FakeNewsNet GossipCop
Benchmarks: FakeNewsNet PolitiFact FakeNewsNet GossipCop
Metrics: accuracy precision recall F1
Baselines: TruthFinder UFD GTUT UFNDA (UMD)^2 SAFE EANN BiGCN GACL
Builds on: Graph Isomorphism Network graph autoencoder GraphMAE UPFD
Closest papers by meaning
- Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference (AAAI 2025 · similarity 0.86)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.85)
- Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.85)
- Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection (AAAI 2025 · similarity 0.83)
- Propagation Tree Is Not Deep: Adaptive Graph Contrastive Learning Approach for Rumor Detection (AAAI 2024 · similarity 0.83)
- Collaborative Evolution: Multi-Round Learning Between Large and Small Language Models for Emergent Fake News Detection (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.