Research map / Graph learning / Graph self-supervised learning
Single-View Graph Contrastive Learning with Soft Neighborhood Awareness
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 SIGNA, a single-view graph contrastive learning framework that uses dropout-perturbed embeddings, stochastic neighbor masking (neighbors randomly flip between positive and negative), and a normalized JSD estimator, achieving strong node-level results with MLP encoders on transductive tasks.
Problem
Graph contrastive learning typically relies on cross-view contrast, which brings augmentation design effort, information loss between views, and high computational cost; naively pulling all neighbors together overfits topology given the inconsistency between structural connection and semantic simila
Approach
Single-branch contrast within one view: dropout yields randomly noised neighbor embedding pairs, a Bernoulli mask switches a fraction of neighbors between positive and negative each epoch (soft neighborhood awareness), and a cosine-based normalized JSD discriminator is used as the loss.
Key results
- On PPI, SIGNA reaches 92.25 F1 versus 70.49 for the best unsupervised baseline (BGRL-GAT-Encoder), a 21.74% margin.
- On transductive node classification, SIGNA with an MLP encoder improves over the runner-up by 1.55% (Wiki CS), 2.04% (Amazon Photo), 0.58% (Amazon Computers), 1.71% (Coauthor CS), and 0.66% (Coauthor Physics).
- MLP-based encoder inference is 109x to 331x faster than GCN-based encoder inference across the five transductive datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SIGNA instead of the strongest unsupervised baseline | Unsupervised node classification · Wiki CS; Amazon Photo; Amazon Computers; Coauthor CS; Coauthor Physics | accuracy: 1.55%; 2.04%; 0.58%; 1.71%; 0.66% | Helped |
| Use SIGNA instead of the previous best method | Inductive node classification · Flickr | micro-averaged F1-score: 3.13% | Helped |
| Use SIGNA instead of the previous best method | Inductive node classification · PPI | micro-averaged F1-score: 21.74% | Helped |
| Use SIGNA instead of AFGRL | Node clustering · Wiki CS, Amazon Photo, Amazon Computers, Coauthor CS, Coauthor Physics | NMI and Homogeneity | Helped |
| Add dropout to the encoder (ablation) | SIGNA · Transductive node classification (five datasets) | accuracy: 80.04$\pm$0.55 vs 80.91$\pm$0.46 (Wiki CS); 93.85$\pm$0.26 vs 95.32$\pm$0.19 (Amazon Photo); 89.72$\pm$0.36 vs 90.46$\pm | Helped |
| Use dropout rather than node feature masking (ablation) | SIGNA · Transductive node classification (five datasets) | accuracy: 75.21$\pm$0.45 vs 80.91$\pm$0.46 (Wiki CS); 91.16$\pm$0.53 vs 95.32$\pm$0.19 (Amazon Photo); 87.43$\pm$0.36 vs 90.46$\pm | Helped |
| Add stochastic neighbor masking rather than treating all neighbors as positive (ablation) | SIGNA · Transductive node classification (five datasets) | accuracy: 80.58$\pm$0.65 vs 80.91$\pm$0.46 (Wiki CS); 94.87$\pm$0.28 vs 95.32$\pm$0.19 (Amazon Photo); 90.09$\pm$0.34 vs 90.46$\pm | Helped |
| Keep some neighbors unmasked rather than mask all neighbors (ablation) | SIGNA · Transductive node classification (five datasets) | accuracy: 68.14$\pm$0.69 vs 80.91$\pm$0.46 (Wiki CS); 80.59$\pm$0.99 vs 95.32$\pm$0.19 (Amazon Photo); 76.64$\pm$0.60 vs 90.46$\pm | Helped |
| Use Norm-JSD rather than JSD (ablation) | SIGNA · Transductive node classification (five datasets) | accuracy: 78.11$\pm$0.59 vs 80.91$\pm$0.46 (Wiki CS); 76.90$\pm$0.54 vs 95.32$\pm$0.19 (Amazon Photo); 70.02$\pm$0.77 vs 90.46$\pm | Helped |
| Use Norm-JSD rather than InfoNCE (ablation) | SIGNA · Transductive node classification (five datasets) | accuracy: 79.03$\pm$0.71 vs 80.91$\pm$0.46 (Wiki CS); 93.64$\pm$0.39 vs 95.32$\pm$0.19 (Amazon Photo); 88.10$\pm$0.35 vs 90.46$\pm | Helped |
| Use an MLP instead of a GCN as the base encoder | SIGNA · Transductive node classification (five datasets) | inference time: $151\times$; $109\times$; $196\times$; $180\times$; $331\times$ | Helped |
Surprising findings
- Replacing the GCN encoder with an MLP encoder in transductive tasks still improves accuracy, since soft neighborhood awareness supplies structural signal through the loss rather than message passing.
- Masking all neighbors (pure feature-based contrast with no structural positives) drops Wiki CS accuracy from 80.91 to 68.14, while removing all three components drops Coauthor Physics to 50.57.
Limitations
- Sensitive to dropout rate p, especially above 0.5, and hyperparameters are grid-searched per dataset.
- Authors note the MLP encoder is used only for transductive tasks; inductive tasks require graph convolution layers.
Experimental setup
Datasets: Wiki CS Amazon Photo Amazon Computers Coauthor CS Coauthor Physics Flickr PPI
Benchmarks: Wiki CS Amazon Photo Amazon Computers Coauthor CS Coauthor Physics Flickr PPI
Metrics: micro-averaged F1-score NMI Homogeneity inference time
Baselines: GAE VGAE DGI GMI MVGRL GRACE CCA-SSG SUGRL BGRL AFGRL GCA Subg-Con
Builds on: GRACE DGI GraphSAGE InfoNCE Deep Graph Infomax
Closest papers by meaning
- Neighbor Contrastive Learning on Learnable Graph Augmentation (AAAI 2023 · similarity 0.90)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.90)
- Return to Basics: Very Simple Graph Contrastive Learning via Noise Cancellation Principle (NeurIPS 2026 · similarity 0.89)
- CL-GCL: Comprehensive and Lightweight Graph Contrastive Learning (ICML 2026 · similarity 0.89)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.88)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · 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.