Research map / Graph learning / Graph self-supervised learning
SOLA-GCL: Subgraph-Oriented Learnable Augmentation Method for 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 SOLA-GCL, a graph contrastive learning method that partitions graphs into subgraphs and learns end-to-end augmentation policies (node dropping, feature masking, intra/inter-edge perturbation, subgraph swapping) to generate graph views.
Problem
Existing graph contrastive learning methods neglect intra-subgraph characteristics and inter-subgraph relationships, and either use random augmentations that alter semantics or uniform learnable augmentations that adapt poorly across datasets.
Approach
Partition each graph into densely connected subgraphs (Louvain or RDKit), then use a learnable graph view generator with a Gumbel-Softmax subgraph augmentation selector, per-subgraph view generators, and a differentiable view assembler to produce subgraph-aware contrastive views.
Key results
- Unsupervised learning: SOLA-GCL reports 90.49% on MUTAG and 82.07% on NCI1, best among compared methods in the unsupervised table.
- Transfer learning: SOLA-GCL reports 74.08 on BBBP, 81.49 on ClinTox, and 78.68 on HIV, the best among compared methods, while trailing AD-GCL on SIDER and AutoGCL on BACE.
- Semi-supervised (10% labels): SOLA-GCL reports 77.63 on PROTEINS and 79.43 on DD, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SOLA-GCL instead of the strongest listed baseline | graph contrastive learning · Unsupervised classification: MUTAG, NCI1, COLLAB, IMDB-B, IMDB-M, Facebook, DBLP | accuracy: 90.49$\pm$2.22 vs 89.68$\pm$0.57; 82.07$\pm$1.32 vs 82.00$\pm$0.29; 73.80$\pm$1.32 vs 73.48$\pm$0.83; 74.20$\pm$1.89 vs | Helped |
| use SOLA-GCL instead of the strongest listed baseline | graph contrastive learning · Unsupervised Infections | accuracy: 53.50$\pm$1.92 vs 53.50$\pm$1.38 | No clear effect |
| use SOLA-GCL instead of the strongest listed baseline | graph contrastive learning · Transfer BBBP, ToxCast, ClinTox, HIV | accuracy: 74.08$\pm$0.82 vs 73.36$\pm$0.77; 64.50$\pm$0.41 vs 63.47$\pm$0.38; 81.49$\pm$1.31 vs 80.99$\pm$3.38; 78.68$\pm$1.01 vs | Helped |
| use SOLA-GCL instead of the strongest listed baseline | graph contrastive learning · Transfer SIDER and BACE | accuracy: 62.79$\pm$0.91 vs 63.28$\pm$0.79; 82.64$\pm$1.01 vs 83.26$\pm$1.13 | Hurt |
| use SOLA-GCL instead of the strongest listed baseline | graph contrastive learning · Semi-supervised PROTEINS (10% Data), DD (10% Data), IMDB-B (10% Data) | accuracy: 77.63$\pm$3.10 vs 75.76$\pm$0.52; 79.43$\pm$4.31 vs 78.89$\pm$0.18; 74.50$\pm$4.65 vs 73.70$\pm$4.88 | Helped |
| use SOLA-GCL instead of the strongest listed baseline | graph contrastive learning · Semi-supervised NCI1 and COLLAB (10% Data) | accuracy: 74.94$\pm$1.96 vs 75.18$\pm$0.31; 76.97$\pm$1.31 vs 77.16$\pm$1.48 | Hurt |
| use GN instead of Louvain partitioning (ablation) | SOLA-GCL · MUTAG | accuracy: 90.49$\pm$2.22 -> 89.88$\pm$2.98 | Hurt |
| use GN instead of Louvain partitioning (ablation) | SOLA-GCL · NCI1 | accuracy: 82.07$\pm$1.32 -> 81.56$\pm$2.11 | Hurt |
| use GN instead of Louvain partitioning (ablation) | SOLA-GCL · COLLAB | accuracy: 73.80$\pm$1.32 -> 72.32$\pm$1.88 | Hurt |
| use GN instead of Louvain partitioning (ablation) | SOLA-GCL · IMDB-B | accuracy: 74.20$\pm$1.89 -> 74.10$\pm$2.84 | No clear effect |
Surprising findings
- Ablation shows the Louvain partitioning yields only about 2.75 to 4.91 subgraphs per graph on average, and removing the learnable selector or generator lowers accuracy across datasets.
Limitations
- Edge perturbation uses randomly sampled negative edges to avoid predicting the full adjacency matrix, which is an approximation.
- Relies on graph partition algorithms (Louvain, RDKit) fixed before training, so partition quality bounds the method.
- Does not clearly report hardware or compute resources.
Experimental setup
Datasets: MUTAG NCI1 COLLAB IMDB-B IMDB-M Infections Facebook DBLP PROTEINS DD BBBP ToxCast
Benchmarks: MUTAG NCI1 COLLAB IMDB-B IMDB-M Infections Facebook DBLP PROTEINS DD BBBP ToxCast
Metrics: classification accuracy ROC-AUC
Baselines: GL WL DGK node2vec sub2vec graph2vec InfoGraph GraphCL JOAO JOAOv2 AD-GCL AutoGCL
Builds on: GraphCL JOAO AD-GCL AutoGCL MSSGCL
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 0.90)
- Community-Invariant Graph Contrastive Learning (ICML 2024 · similarity 0.89)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.89)
- Unified Graph Augmentations for Generalized Contrastive Learning on Graphs (NeurIPS 2024 · similarity 0.89)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.89)
- Graph Contrastive Learning with Joint Spectral Augmentation of Attribute and Topology (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.