Research map / Graph learning / Graph self-supervised learning
Adversarial Contrastive Graph Augmentation with Counterfactual Regularization
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 ACGA, a model-based adversarial graph augmentation method that jointly generates positive (minimal sufficient) and hard negative graphs via a conditional VGAE regularized by counterfactual graph approximations, for self-supervised graph contrastive learning.
Problem
Self-supervised graph contrastive learning relies on rule-based augmentations and model-based methods that only augment positives, neglecting hard negative samples; generating labeled-quality positive/negative pairs without labels is difficult.
Approach
Uses counterfactual graph approximation (via node-degree auxiliary variable and edge counterfactual links) to build typical positive and negative graphs that regularize a Conditional VGAE generator, which adversarially produces positive and hard negative graphs; combined with inter-graph and intra-graph contrastive losses and a theoretical unification of positive/negative augmentation.
Key results
- ACGA reports 84.4% node classification accuracy on Cora vs. 83.4% for the best baseline GAUG, and 74.2% on CiteSeer vs. 72.3% for GAUG.
- On link prediction, ACGA achieves 94.7 AUC on Cora vs. 93.6 for JOAO and 93.0 for GCA-EV.
- Ablation on Cora shows removing counterfactual regularization (NO REG) drops accuracy from 84.4 to 81.9, and negative-only augmentation (N-AGA) gives 71.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ACGA instead of GAUG | Node classification · Cora | accuracy: 83.4±0.4 -> 84.4±0.6 | Helped |
| Use ACGA instead of MVGRL | Node classification · BlogCatalog | accuracy: 79.7±0.4 -> 77.5±0.8 | Hurt |
| Use ACGA instead of JOAO | Link prediction · Cora | AUC: 93.6±0.3 -> 94.7±0.2 | Helped |
| Use ACGA instead of GCA-EV | Link prediction · BlogCatalog | AP: 86.2±0.5 -> 86.0±0.2 | No clear effect |
| Use N-AGA alone instead of the full ACGA (ablation) | ACGA node classification · Cora | accuracy: 84.4±0.6 -> 71.3±8.3 | Hurt |
| Use P-AGA alone instead of the full ACGA (ablation) | ACGA node classification · Cora | accuracy: 84.4±0.6 -> 82.6±0.7 | Hurt |
| Replace typical graphs with random graphs (ablation) | ACGA node classification · Cora | accuracy: 84.4±0.6 -> 82.1±0.8 | Hurt |
| Remove regularization (ablation) | ACGA node classification · Cora | accuracy: 84.4±0.6 -> 81.9±0.9 | Hurt |
| Remove intra-GCL (ablation) | ACGA node classification · Cora | accuracy: 84.4±0.6 -> 82.9±0.8 | Hurt |
| Remove KL loss (ablation) | ACGA node classification · Cora | accuracy: 84.4±0.6 -> 83.5±0.7 | Hurt |
Surprising findings
- Negative-only adversarial augmentation (N-AGA) alone performs much worse than positive-only (P-AGA), e.g., 71.3 vs. 82.6 on Cora, and has high variance (±8.3).
Limitations
- Large datasets (PubMed, Coauthor-CS) caused OOM for most baselines, requiring sampling of 60% of nodes for graphs over 10k nodes.
- Counterfactual graph approximation has O(N^3) time complexity, which may be costly for large graphs.
- Evaluated only on node classification and link prediction; graph classification left for future work.
Experimental setup
Datasets: Cora CiteSeer PubMed BlogCatalog Flickr Air-USA Coauthor-CS
Benchmarks: Cora CiteSeer PubMed BlogCatalog Flickr Air-USA Coauthor-CS
Metrics: accuracy AUC AP
Baselines: AD-GCL JOAO GAUG MVGRL GCA AFGRL GAE GCN
Builds on: AD-GCL GAUG VGAE GCA Graph Information Bottleneck
Closest papers by meaning
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.89)
- Improving Graph Contrastive Learning via Adaptive Positive Sampling (CVPR 2024 · similarity 0.89)
- Community-Invariant Graph Contrastive Learning (ICML 2024 · similarity 0.88)
- Unified Graph Augmentations for Generalized Contrastive Learning on Graphs (NeurIPS 2024 · similarity 0.88)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.88)
- Rethinking Graph Masked Autoencoders through Alignment and Uniformity (AAAI 2024 · similarity 0.88)
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.