Research map / Graph learning / Graph self-supervised learning
A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning
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.
Identifies an implicit Embedding and Ignoring Conflict (EIC) in InfoNCE-based graph contrastive learning and proposes PiGCL, which dynamically ignores conflicting negative pairs detected via similarity-change gradients and compensates the loss with a scaling factor.
Problem
InfoNCE-based graph contrastive learning suffers from an implicit conflict between GNN message passing and the InfoNCE negative-pair objective, causing node representations of negative pairs to remain overly similar.
Approach
Introduces the Embedding and Ignoring Conflict (EIC) concept, detects conflicting negative pairs by tracking per-iteration changes in their similarity as a gradient proxy, temporarily ignores the top-r% of such pairs in the loss, and rescales the remaining negative terms (alpha) plus a compensation term (phi) to keep gradient magnitude stable.
Key results
- PiGCL reaches 84.63% on Cora node classification, outperforming GRACE by about 1.43% average F1 across datasets.
- On node clustering, PiGCL improves over GRACE by an average of 5.44% NMI and 7.02% ARI.
- Parameter analysis shows accuracy with ignore rate r>0 generally beats r=0, with a unimodal dependence on r.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PiGCL instead of GCA | InfoNCE-based graph contrastive learning · Cora | F1-score: 83.42 ± 0.78 -> 84.63 ± 0.78 | Helped |
| Use PiGCL instead of ProGCL | InfoNCE-based graph contrastive learning · CiteSeer | F1-score: 72.85 ± 0.92 -> 73.51 ± 0.64 | Helped |
| Use PiGCL instead of GCA | InfoNCE-based graph contrastive learning · PubMed | F1-score: 86.12 ± 0.22 -> 86.75 ± 0.20 | Helped |
| Use PiGCL instead of Supervised GCN | node classification · CS | F1-score: 93.03 -> 93.30 ± 0.09 | Helped |
| Use PiGCL instead of Local-GCL | InfoNCE-based graph contrastive learning · Photo | F1-score: 92.99 ± 0.34 -> 93.14 ± 0.30 | Helped |
| Use PiGCL instead of Local-GCL | InfoNCE-based graph contrastive learning · Computers | F1-score: 88.64 ± 0.25 -> 89.25 ± 0.27 | Helped |
| Use PiGCL instead of Local-GCL | graph contrastive learning · Cora | NMI: 0.5386 -> 0.5494 | Helped |
| Use PiGCL instead of Local-GCL | graph contrastive learning · Cora | ARI: 0.4479 -> 0.4670 | Helped |
| Use PiGCL instead of Local-GCL | graph contrastive learning · CiteSeer | NMI: 0.4508 -> 0.4581 | Helped |
| Use PiGCL instead of Local-GCL | graph contrastive learning · CiteSeer | ARI: 0.4494 -> 0.4720 | Helped |
| Use PiGCL instead of GRACE | graph contrastive learning · PubMed | NMI: 0.3508 -> 0.3784 | Helped |
| Use PiGCL instead of Local-GCL | graph contrastive learning · PubMed | ARI: 0.3304 -> 0.3612 | Helped |
Surprising findings
- GRACE, which uses inter-node negative relations, does not clearly outperform BGRL, DGI and GBT on node classification, which the authors attribute to EIC.
Limitations
- Theoretical proof of EIC is given for a single-layer GCN and is argued to generalize only informally to other GNNs.
- Performance is sensitive to the ignore ratio r; too large a ratio degrades accuracy.
Experimental setup
Datasets: Cora CiteSeer PubMed Amazon Photo Amazon Computers Coauthor CS
Benchmarks: Cora CiteSeer PubMed Amazon Photo Amazon Computers Coauthor CS
Metrics: F1-score NMI ARI
Baselines: GRACE GCA Local-GCL ProGCL BGRL MVGRL DGI GBT Node2vec DeepWalk GCN
Builds on: GRACE InfoNCE GCA BGRL
Closest papers by meaning
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.93)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.92)
- Return to Basics: Very Simple Graph Contrastive Learning via Noise Cancellation Principle (NeurIPS 2026 · similarity 0.92)
- An Empirical Study of Graph Contrastive Learning (NeurIPS 2021 · similarity 0.91)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.90)
- Beyond Homophily: Graph Contrastive Learning with Macro-Micro Message Passing (AAAI 2025 · similarity 0.90)
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.