Research map / Graph learning / Graph self-supervised learning
LOHA: Direct Graph Spectral Contrastive Learning Between Low-Pass and High-Pass Views
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 LOHA, an augmentation-free spectral graph contrastive framework that treats low-pass and high-pass filtered views as negatives for each other while reuniting node features via a spectral signal trend composite feature.
Problem
Self-supervised graph contrastive learning with spectral GNNs struggles to exploit low-pass and high-pass views across graphs with varying homophily; contrasting the two views can separate a node's features, so a consistent node identity must be preserved.
Approach
Directly contrast low-pass and high-pass views (treating them as mutual negatives) using learnable sliding cosine-parameterized Chebyshev filters, then reunite each node's features with a composite loss based on a "spectral signal trend" feature that is less sensitive to filter changes.
Key results
- LOHA achieves average accuracy 77.1% across nine datasets, versus 74.3% for the runner-up PolyGCL and 74.4% for the best fully-supervised ChebNetII.
- On Cornell, Texas, and Wisconsin, LOHA reaches 90.7%, 93.4%, and 95.9% accuracy, surpassing fully-supervised baselines.
- Replacing PolyGCL's loss with the LOHA loss improves Cornell from 82.36% to 86.73% and Wisconsin from 84.03% to 91.06%, but lowers Cora from 87.57% to 86.21%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LOHA rather than the runner-up baseline | Node classification · Cornell | accuracy | Helped |
| Use LOHA rather than the runner-up baseline | Node classification · Texas | accuracy | Helped |
| Use LOHA rather than the runner-up baseline | Node classification · Wisconsin | accuracy | Helped |
| Use the LOHA-var-3 spectral signal trend variant instead of LOHA (ablation) | LOHA · Nine real-world datasets | accuracy | Hurt |
| Use the LOHA-var-1 spectral signal trend variant instead of LOHA (ablation) | LOHA · Nine real-world datasets | accuracy | Hurt |
| Remove the sliding interpolation trick (ablation) | LOHA · Real-world graph datasets | accuracy | Hurt |
| Remove the spectral-feature reunion loss (ablation) | LOHA · Real-world graph datasets | accuracy | Hurt |
| Remove the low-pass/high-pass contrastive term (ablation) | LOHA · Real-world graph datasets | accuracy | Hurt |
| Replace PolyGCL's loss with LOHA loss (ablation) | PolyGCL · Cornell | accuracy: 82.36±3.42 -> 86.73±2.34 | Helped |
| Replace PolyGCL's loss with LOHA loss (ablation) | PolyGCL · Texas | accuracy: 87.92±1.46 -> 91.51±0.92 | Helped |
| Replace PolyGCL's loss with LOHA loss (ablation) | PolyGCL · Cora | accuracy: 87.57±0.73 -> 86.21±0.65 | Hurt |
| Replace PolyGCL's loss with LOHA loss (ablation) | PolyGCL · PubMed | accuracy: 86.91±0.34 -> 86.96±0.32 | No clear effect |
Surprising findings
- Band-stop and band-pass filter variants achieve results competitive with low/high-pass filter sets in the PolyGCL framework, questioning whether the gains come from the low/high-pass specialties.
- LOHA underperforms the PolyGCL baseline on homophilic Cora and Citeseer (85.1% and 77.1% vs. 87.6% and 78.9%).
Limitations
- Performance on homophilic graphs Cora and Citeseer is below PolyGCL.
- Theoretical analysis relies on assumptions (independence across dimensions, bounded features) that may not hold in practice.
Experimental setup
Datasets: Cora Citeseer PubMed Cornell Texas Wisconsin Actor Chameleon Amazon-ratings
Benchmarks: Cora Citeseer PubMed Cornell Texas Wisconsin Actor Chameleon Amazon-ratings
Metrics: node classification accuracy
Baselines: MLP GCN ChebNet BernNet ChebNetII DGI GraphCL GRACE BGRL GREET GraphACL PolyGCL
Builds on: PolyGCL GRACE ChebNetII BernNet InfoNCE
Closest papers by meaning
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.89)
- SLOG: An Inductive Spectral Graph Neural Network Beyond Polynomial Filter (ICML 2024 · similarity 0.89)
- Unifying Homophily and Heterophily for Spectral Graph Neural Networks via Triple Filter Ensembles (NeurIPS 2024 · similarity 0.88)
- Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.88)
- HoloNets: Spectral Convolutions do extend to Directed Graphs (ICLR 2024 · similarity 0.88)
- HeroFilter: Adaptive Spectral Graph Filter for Varying Heterophilic Relations (NeurIPS 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.