Research map / Graph learning / Graph self-supervised learning

Self-supervised Masked Graph Autoencoder via Structure-aware Curriculum

ICML 2025 · Graph learning › Graph self-supervised learning · approach: Heterophily and imbalance on graphs

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 Cur-MGAE, a masked graph autoencoder that masks edges in an easy-to-hard curriculum, using reconstruction residuals as difficulty scores and a self-paced scheduler, evaluated on node classification and link prediction.

Problem

Masked graph autoencoders treat all masked edges uniformly, ignoring that edges differ in reconstruction difficulty, which yields suboptimal node representations for graph self-supervised learning.

Approach

Score each edge's reconstruction difficulty by its structural residual from the current model, then use a self-paced mask scheduler (Lagrangian-relaxed selection with a growing lambda and a random/difficulty split ratio) to mask edges from easy to hard within a masked-autoencoder framework with a cross-correlation decoder.

Key results

• Node classification: Cur-MGAE reaches 85.86% on Pubmed (vs. 84.19 for S2GAE, the strongest baseline by reported margin) and 73.00% on OGBN-arxiv, with the best average rank (2.83) among generative and contrastive baselines.
• Link prediction: Cur-MGAE achieves 68.50 Hits@20 on OGBL-ddi and 5.96 Hits@10 on OGBL-ppa, roughly 2% above the strongest baselines on these datasets.
• Ablation: removing the curriculum masking drops Cora node classification from 87.25 to 86.08 and OGBL-ddi link prediction from 68.50 to 61.70.

Measured effects reported in the paper

Surprising findings

• Ablation with split ratio 0 (purely random selection among masked edges) gives a higher Cora link prediction AUC (94.12) than the full model's split-ratio variant at 1 (93.80), but the full model still scores best overall (95.22).

Limitations

• Gains over baselines are often small relative to standard deviations (e.g., Cora node classification 87.25 ± 0.55 vs MaskGAE 87.31 ± 0.05), and the method does not win on all datasets.
• Theoretical guarantees (saddle point avoidance, second-order convergence) rely on smoothness and KL assumptions and do not directly show improved representation quality.
• Many hyperparameters (split ratio, mask ratio, lambda_initial, dropout, decoder hidden size) are tuned per dataset, and the sensitivity analysis is limited to Cora and PubMed.

Experimental setup

Datasets: Cora Citeseer Pubmed Coauthor-CS Coauthor-Physics OGBN-arxiv OGBL-ddi OGBL-collab OGBL-ppa synthetic multi-Gaussian graphs with homophily 0.1/0.5/0.9

Benchmarks: Cora Citeseer Pubmed Coauthor-CS Coauthor-Physics OGBN-arxiv OGBL-ddi OGBL-collab OGBL-ppa

Metrics: accuracy AUC Hits@N average rank

Baselines: DGI GIC MVGRL BGRL GAE GraphSAGE ARGVA GPT-GNN RRL GraphMAE GraphMAE2 MaskGAE

Builds on: S2GAE GraphMAE MaskGAE self-paced learning Zhang et al. 2023a self-paced graph structure learning

Closest papers by meaning

• Teacher-guided Edge Discriminator for Personalized Graph Masked Autoencoder (AAAI 2025 · similarity 0.92)
• Rethinking Graph Masked Autoencoders through Alignment and Uniformity (AAAI 2024 · similarity 0.90)
• Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.88)
• Self-Guided Masked Autoencoders for Domain-Agnostic Self-Supervised Learning (ICLR 2024 · similarity 0.88)
• The quest for the GRAph Level autoEncoder (GRALE) (NeurIPS 2025 · similarity 0.87)
• GraphCroc: Cross-Correlation Autoencoder for Graph Structural Reconstruction (NeurIPS 2024 · similarity 0.86)