Research map / Graph learning / Graph self-supervised learning
S3GCL: Spectral, Swift, Spatial 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.
Proposes S3GCL, a label-free graph contrastive learning framework that uses cosine-parameterized Chebyshev low/high-pass filters on a GNN branch and a spatially-augmented MLP encoder trained with a cross-pass objective, enabling augmentation-free training and fast MLP inference.
Problem
Graph contrastive learning mostly assumes homophily (relying on low-pass filtering), generalizes poorly to heterophilic graphs, and relies on neighborhood aggregation at inference, which is slow for large or latency-sensitive graphs.
Approach
Learnable cosine-parameterized Chebyshev polynomial filters produce low-pass and high-pass biased views; an MLP encoder (full-pass) is contrasted against these GNN-filtered views, with structural neighbors as positives for the low-pass part and feature-kNN neighbors as positives for the high-pass part, removing the need for graph augmentation.
Key results
- S3GCL reports 88.47% on Cora and 76.31% on CiteSeer, and ranks first on 11 of 14 datasets in the reported comparison.
- On Ogbn-Arxiv, S3GCL inference takes 2.23 ms versus 389.6 ms for a 2-layer BGRL, a reported 174x speedup.
- Ablations show removing structural or semantic spatial positive pairs lowers accuracy, e.g. Roman-empire drops from 66.27 to 63.11 without semantic pairs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace cosine-parameterized Chebyshev filters with linear parameterization (ablation) | S3GCL · Cora | accuracy (%): -0.66 (88.47 to 87.81) | Hurt |
| Remove cosine parameterization (plain ChebNet interpolation) (ablation) | S3GCL · Cora | accuracy (%): -4.14 (88.47 to 84.33) | Hurt |
| Remove structural neighbor positive pairs (full-low part) (ablation) | S3GCL · Cora | accuracy (%): -1.60 (88.47 to 86.87) | Hurt |
| Remove semantic neighbor positive pairs (full-high part) (ablation) | S3GCL · Cora | accuracy (%): -0.93 (88.47 to 87.54) | Hurt |
| Replace spatial-pair cross-pass objective with plain InfoNCE (no spatial pairs) (ablation) | S3GCL · Cora | accuracy (%): -3.50 (88.47 to 84.97) | Hurt |
| Use MLP-based cross-architecture GCN-MLP instead of full S3GCL filters (ablation) | S3GCL · Cora | accuracy (%): -3.15 (88.47 to 85.32) | Hurt |
| Replace cosine-parameterized Chebyshev with linear param. / remove param. on PubMed (ablation) | S3GCL · PubMed | accuracy (%): -4.00 to -0.92 across variants (vs 87.89) | Hurt |
| Adding spatial neighbor positive pairs and cosine Chebyshev filters on Roman-empire (ablation) | S3GCL · Roman-empire | accuracy (%): +1.51 over Linear Param. (66.27 vs 65.77) | Helped |
| Use S3GCL (MLP encoder, cross-pass objective) instead of prior GCL methods for inference | MLP encoder · Obgn-Arxiv | inference time (ms): 389.6ms to 2.23ms (about 174x) | Helped |
| Use S3GCL instead of prior GCL methods on the 14-dataset benchmark | S3GCL · Table 1 (14 datasets) | accuracy (%): best on 11 of 14 datasets | Helped |
| Replace ChebNet-style fixed filters with cosine-parameterized filter on Cora | S3GCL · Cora | accuracy (%): +0.8 over next best (88.47 vs 87.67) | Helped |
Limitations
- Training still requires polynomial graph filters, so training cost is not reduced; efficiency gains only at inference.
- Filter parameters are not learned with labels, and performance depends on hyperparameters such as alpha and k.
Experimental setup
Datasets: Cora CiteSeer PubMed Amz-Comp Amz-Photo Coauthor-CS Obgn-Arxiv Cornell Texas Wisconsin Actor Roman-empire
Benchmarks: Cora CiteSeer PubMed Amz-Comp Amz-Photo Coauthor-CS Obgn-Arxiv Cornell Texas Wisconsin Actor Roman-empire
Metrics: node classification accuracy inference time
Baselines: DGI GMI MVGRL GRACE GBT BGRL GREET DSSL SP-GCL
Builds on: GRACE ChebNet BGRL InfoNCE ChebNetII
Closest papers by meaning
- PolyGCL: GRAPH CONTRASTIVE LEARNING via Learnable Spectral Polynomial Filters (ICLR 2024 · similarity 0.94)
- Beyond Homophily: Graph Contrastive Learning with Macro-Micro Message Passing (AAAI 2025 · similarity 0.92)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.91)
- Efficient Contrastive Learning for Fast and Accurate Inference on Graphs (ICML 2024 · similarity 0.91)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.91)
- Improving Graph Contrastive Learning via Adaptive Positive Sampling (CVPR 2024 · 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.