Research map / Graph learning / Graph self-supervised learning
Improving Graph Contrastive Learning via Adaptive Positive Sampling
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 HEATS, a graph contrastive learning framework that builds block-diagonal, idempotent affinity matrices via a self-expressive objective to select global positive samples, alternating with contrastive optimization.
Problem
Existing graph contrastive learning positive sampling is incomplete (limited to local neighborhoods) and blind (no criterion beyond neighbors), so GCLs perform suboptimally on both homophilic and heterophilic graphs.
Approach
Motivated by the observation that the affinity matrix of the optimal positive sample set is block-diagonal with equal weights per block, learn such an affinity matrix via a self-expressive objective with block (idempotent, k-block) constraints, and alternate its construction with contrastive optimization over the GRACE framework.
Key results
- On Cora, HEATS reaches 87.10% accuracy vs. 84.79% for GRACE; on CiteSeer it improves over GRACE by 2.32 points.
- On heterophilic graphs HEATS improves over GRACE by 12.11%, 7.29%, and 15.88% on Cornell, Texas, and Wisconsin.
- On CIFAR-100 with ResNet-50, SimCLR+HEATS improves from 49.31% to 53.81%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HEATS adaptive positive sampling instead of MVGRL | Unsupervised node classification · Cora | accuracy: 86.77 ± 0.33 -> 87.10 ± 1.40 | Helped |
| Use HEATS adaptive positive sampling instead of MVGRL | Unsupervised node classification · CiteSeer | accuracy: 73.71 ± 0.48 -> 75.26 ± 1.25 | Helped |
| Use HEATS adaptive positive sampling instead of GCA | Unsupervised node classification · PubMed | accuracy: 85.22 ± 0.73 -> 85.41 ± 0.82 | Helped |
| Use HEATS adaptive positive sampling instead of GCA | Unsupervised node classification · Wiki-CS | accuracy: 79.35 ± 0.12 -> 79.99 ± 1.59 | Helped |
| Use HEATS adaptive positive sampling instead of BGRL | Unsupervised node classification · Computers | accuracy: 88.92 ± 0.33 -> 89.29 ± 1.12 | Helped |
| Use HEATS adaptive positive sampling instead of HomoGCL | Unsupervised node classification · Photo | accuracy: 93.59 ± 0.27 -> 94.65 ± 1.71 | Helped |
| Use HEATS adaptive positive sampling instead of the strongest listed baseline | Unsupervised node classification · Cornell | accuracy: 64.30 ± 5.43 -> 66.97 ± 6.74 | Helped |
| Use HEATS adaptive positive sampling instead of JKNet | Unsupervised node classification · Texas | accuracy: 65.35 ± 4.86 -> 64.86 ± 4.68 | Hurt |
| Use HEATS adaptive positive sampling instead of MVGRL | Unsupervised node classification · Wisconsin | accuracy: 62.37 ± 4.32 -> 65.88 ± 5.56 | Helped |
| Use HEATS adaptive positive sampling instead of MVGRL | Unsupervised node classification · Chameleon | accuracy: 51.07 ± 2.68 -> 49.96 ± 1.86 | Hurt |
| Use HEATS adaptive positive sampling instead of HomoGCL | Unsupervised node classification · Squirrel | accuracy: 38.71 ± 0.85 -> 36.24 ± 1.11 | Hurt |
| Use HEATS adaptive positive sampling instead of SELENE | Unsupervised node classification · Actor | accuracy: 30.12 ± 0.76 -> 29.91 ± 1.16 | Hurt |
Surprising findings
- HEATS does not beat all unsupervised baselines on Chameleon, Squirrel, and Actor, where MVGRL and other methods remain stronger.
Limitations
- Scalability to large datasets is identified as future work.
- Gains on some heterophilic datasets (Chameleon, Squirrel, Actor) are not superior to baselines.
- Results on high-variance datasets such as Cornell and Texas have large standard deviations.
Experimental setup
Datasets: Cora CiteSeer PubMed Wiki-CS Computers Photo Cornell Texas Wisconsin Chameleon Squirrel Actor
Benchmarks: Cora CiteSeer PubMed Wiki-CS Computers Photo Cornell Texas Wisconsin Chameleon Squirrel Actor
Metrics: classification accuracy
Baselines: GCN GAT JKNet DeepWalk Node2Vec GAE VGAE DGI MVGRL GRACE GCA BGRL
Builds on: GRACE SimCLR HomoGCL subspace clustering self-expressive model
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 0.92)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.90)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.90)
- Edge Contrastive Learning: An Augmentation-Free Graph Contrastive Learning Model (AAAI 2025 · similarity 0.89)
- Beyond Homophily: Graph Contrastive Learning with Macro-Micro Message Passing (AAAI 2025 · similarity 0.89)
- Adversarial Contrastive Graph Augmentation with Counterfactual Regularization (AAAI 2025 · similarity 0.89)
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.