Research map / Graph learning / Graph self-supervised learning
Graph Contrastive Learning with Joint Spectral Augmentation of Attribute and Topology
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 GCL-JAM, which turns node attributes into attribute nodes to form an attribute-interpolated graph and applies spectral augmentation jointly to attributes and topology for graph contrastive learning.
Problem
Existing graph augmentation strategies for graph contrastive learning perturb node attributes and graph topology separately, ignoring their correlation, which limits representation quality on both homophilic and heterophilic graphs.
Approach
Abstract node attributes as attribute nodes to build an attribute-interpolated heterogeneous graph, then perform spectral augmentation (perturbing edges with the largest Laplacian eigenvalue change) on it so attributes and topology are augmented jointly; the authors prove this links GNN objectives to spectral clustering and perturbs more inter-class edges and noise attributes.
Key results
- GCL-JAM achieves 85.54% on Cora, 82.09% on Wiki-CS and 94.35% on Photo, best on five of six homophilic datasets.
- On heterophilic graphs GCL-JAM reaches 66.37% on Chameleon and 49.84% on Squirrel, beating the second-best method by 6.74% and 12.76%.
- Plugging JAM augmentation into GRACE, MVGRL and BGRL yields consistent gains, e.g., GRACE on Chameleon rises from 48.05% to 66.37%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| jointly spectrally augment topology and attributes rather than topology alone (ablation) | GCL-JAM · CiteSeer, Wiki-CS, Chameleon, Squirrel | accuracy (%): 72.79±0.61 -> 73.29±0.39; 81.36±0.13 -> 82.09±0.42; 51.26±1.40 -> 66.37±2.37; 35.23±1.09 -> 49.84±0.93 | Helped |
| jointly spectrally augment topology and attributes rather than attributes alone (ablation) | GCL-JAM · CiteSeer, Wiki-CS, Chameleon, Squirrel | accuracy (%): 71.80±0.75 -> 73.29±0.39; 78.84±0.29 -> 82.09±0.42; 55.80±2.15 -> 66.37±2.37; 34.96±1.27 -> 49.84±0.93 | Helped |
| replace spectral augmentation with random spatial augmentation (ablation) | GCL-JAM · CiteSeer, Wiki-CS, Chameleon, Squirrel | accuracy (%): 71.06±0.45 -> 73.29±0.39; 78.61±0.23 -> 82.09±0.42; 49.65±1.79 -> 66.37±2.37; 32.46±1.18 -> 49.84±0.93 | Hurt |
| add JAM augmentation to GRACE (ablation) | GRACE · Cora, Photo, Chameleon, Squirrel | accuracy (%): 83.44±0.39 -> 85.54±0.53; 92.65±0.32 -> 94.35±0.21; 48.05±1.71 -> 66.37±2.37; 31.33±1.22 -> 49.84±0.93 | Helped |
| add JAM augmentation to MVGRL (ablation) | MVGRL · Cora, Photo, Chameleon, Squirrel | accuracy (%): 83.01±0.42 -> 84.26±0.55; 92.01±0.13 -> 94.26±0.19; 51.07±2.68 -> 64.04±1.90; 35.47±1.29 -> 47.18±0.68 | Helped |
| add JAM augmentation to BGRL (ablation) | BGRL · Cora, Photo, Chameleon, Squirrel | accuracy (%): 82.67±0.78 -> 84.19±0.47; 93.24±0.29 -> 93.81±0.10; 47.46±2.74 -> 69.05±1.18; 32.64±0.78 -> 48.09±0.89 | Mixed |
| use joint attribute-topology spectral augmentation rather than the strongest baseline | GCL-JAM · homophilic graphs: Cora, CiteSeer, Wiki-CS, Computers, Photo | accuracy (%): 85.01±0.89 -> 85.54±0.53; 72.79±0.61 -> 73.29±0.39; 81.36±0.13 -> 82.09±0.42; 89.68±0.35 -> 90.69±0.28; 93.31±0.24 -> 94 | Helped |
| use joint attribute-topology spectral augmentation rather than the strongest baseline | GCL-JAM · heterophilic graphs: Chameleon, Squirrel, Actor, Texas, Wisconsin | accuracy (%): 59.63±2.32 -> 66.37±2.37; 37.08±1.01 -> 49.84±0.93; 30.83±0.77 -> 31.01±0.46; 62.38±5.61 -> 65.83±3.61; 62.37±4.32 -> 64 | Helped |
Surprising findings
- On the Cornell dataset, which has higher homophily, the benefit of attribute nodes is reduced, and GCL-JAM is not best on Actor or Cornell.
Limitations
- Augmentation requires eigenvalue decomposition with cubic cost in N+F, which is expensive for large graphs.
- Gains on Cornell and some heterophilic datasets are smaller or absent (GCL-JAM not best on Actor and Cornell).
Experimental setup
Datasets: Cora CiteSeer PubMed Wiki-CS Computers Photo Chameleon Squirrel Actor Cornell Texas Wisconsin
Benchmarks: Cora CiteSeer PubMed Wiki-CS Computers Photo Chameleon Squirrel Actor Cornell Texas Wisconsin
Metrics: node classification accuracy (%)
Baselines: GCN GAT DeepWalk Node2Vec DGI GMI MVGRL GRACE GCA BGRL GraphMAE SpCo
Builds on: GRACE SPAN-GCL GCN modularity spectral clustering
Closest papers by meaning
- Community-Invariant Graph Contrastive Learning (ICML 2024 · similarity 0.92)
- Unified Graph Augmentations for Generalized Contrastive Learning on Graphs (NeurIPS 2024 · similarity 0.91)
- Hierarchical Topology Isomorphism Expertise Embedded Graph Contrastive Learning (AAAI 2024 · similarity 0.90)
- Edge Contrastive Learning: An Augmentation-Free Graph Contrastive Learning Model (AAAI 2025 · similarity 0.89)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.89)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · 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.