Research map / Graph learning / Graph self-supervised learning
Teacher-guided Edge Discriminator for Personalized Graph Masked Autoencoder
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 TEDMAE, a graph masked autoencoder that separates homophilic and heterophilic edges with a teacher-guided (Node2vec and PCA) edge discriminator and then trains separate encoders/decoders per view with SCE and InfoNCE losses.
Problem
Existing graph masked autoencoders assume homophily and treat homophilic and heterophilic edges uniformly during masking and reconstruction, which hurts representation quality on real graphs that mix both edge types.
Approach
A teacher-guided edge discriminator, supervised by margin ranking losses against anchor node pairs using embeddings from Node2vec (structure) and PCA (features), splits edges into homophilic and heterophilic views; a personalized GMAE then uses separate encoders, masking ratios, and decoders for each view.
Key results
- TEDMAE achieves 84.64% accuracy on Cora and 93.26% on Amazon-Photo, outperforming all listed baselines on those datasets (Table 1).
- On node clustering, TEDMAE achieves the best or runner-up NMI/ARI on 5 of 6 datasets, e.g., 0.7815/0.7638 on CS.
- Ablation on Cora drops from 84.96 to 83.59 when embedding reconstruction is removed, showing reconstruction of embeddings matters most.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the homophilic GMAE module (ablation) | TEDMAE · Cora node classification | accuracy: 83.88 -> 84.96 | Helped |
| Add the homophilic GMAE module (ablation) | TEDMAE · WikiCS node classification | accuracy: 81.06 -> 81.35 | Helped |
| Add the homophilic GMAE module (ablation) | TEDMAE · Amazon-Photo node classification | accuracy: 92.94 -> 93.26 | Helped |
| Add the heterophilic GMAE module (ablation) | TEDMAE · Cora node classification | accuracy: 84.07 -> 84.96 | Helped |
| Add the heterophilic GMAE module (ablation) | TEDMAE · WikiCS node classification | accuracy: 81.31 -> 81.35 | Helped |
| Add the heterophilic GMAE module (ablation) | TEDMAE · Amazon-Photo node classification | accuracy: 92.91 -> 93.26 | Helped |
| Add embedding reconstruction (ablation) | TEDMAE · Cora node classification | accuracy: 83.59 -> 84.96 | Helped |
| Add embedding reconstruction (ablation) | TEDMAE · WikiCS node classification | accuracy: 79.83 -> 81.35 | Helped |
| Add embedding reconstruction (ablation) | TEDMAE · Amazon-Photo node classification | accuracy: 92.29 -> 93.26 | Helped |
| Add feature reconstruction (ablation) | TEDMAE · Cora node classification | accuracy: 84.33 -> 84.96 | Helped |
| Add feature reconstruction (ablation) | TEDMAE · WikiCS node classification | accuracy: 80.54 -> 81.35 | Helped |
| Add feature reconstruction (ablation) | TEDMAE · Amazon-Photo node classification | accuracy: 92.95 -> 93.26 | Helped |
Surprising findings
- TEDMAE underperforms on Amazon-Computers and CoAuthor-Physics, which the authors attribute to high edge homogeneity and low density.
Limitations
- Performance is weaker on sparse or highly homophilic graphs such as Computers and CoAuthor-Physics.
- Hyperparameter search over masking ratios and thresholds adds tuning cost.
- Teacher choice matters; the paper reports GCN/GAT teachers perform well but need labels.
Experimental setup
Datasets: Cora WikiCS Amazon-Computers Amazon-Photo CoAuthor-CS CoAuthor-Physics Cornell Texas Wisconsin Actor
Benchmarks: Cora WikiCS Amazon-Computers Amazon-Photo CoAuthor-CS CoAuthor-Physics Cornell Texas Wisconsin Actor
Metrics: accuracy NMI ARI
Baselines: Deepwalk MVGRL GCA BGRL GREET VGAE GraphMAE GraphMAE2 S2GAE GiGaMAE Bandana MaskGAE
Builds on: GraphMAE GraphMAE2 GiGaMAE GREET Node2vec
Closest papers by meaning
- Self-supervised Masked Graph Autoencoder via Structure-aware Curriculum (ICML 2025 · similarity 0.92)
- Rethinking Graph Masked Autoencoders through Alignment and Uniformity (AAAI 2024 · similarity 0.91)
- The quest for the GRAph Level autoEncoder (GRALE) (NeurIPS 2025 · similarity 0.87)
- Integrating Co-Training with Edge Discrimination to Enhance Graph Neural Networks Under Heterophily (AAAI 2025 · similarity 0.86)
- GraphCroc: Cross-Correlation Autoencoder for Graph Structural Reconstruction (NeurIPS 2024 · similarity 0.86)
- HC-GAE: The Hierarchical Cluster-based Graph Auto-Encoder for Graph Representation Learning (NeurIPS 2024 · similarity 0.86)
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.