Research map / Graph learning / Graph self-supervised learning
Graph Few-Shot Learning via Adaptive Spectrum Experts and Cross-Set Distribution Calibration
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 GRACE, a graph few-shot node classification framework combining a two-expert (low-pass GCN and attention-based high-pass) mixture-of-experts with a KDE-inspired query-guided prototype calibration to handle support-query distribution shift.
Problem
Graph few-shot node classification suffers from fixed global graph filters that ignore heterogeneous local homophily/heterophily, and from support-query distribution shift that makes prototypes unreliable.
Approach
Node-specific gated mixture of a low-pass GCN expert and a high-pass difference-attention expert, plus a kernel-density-weighted calibration that shifts support prototypes toward dense query regions before Euclidean nearest-prototype classification.
Key results
- GRACE reports 66.48% on Cora 2-way 1-shot and 86.19% on Cora 2-way 5-shot, above the strongest listed baselines in its table.
- On ogbn-arxiv 10-way 5-shot, GRACE reaches 55.07% versus 51.11% for TEG.
- Ablation: removing calibration (w/o cal) or either expert lowers accuracy on most datasets, e.g. Cora 2-way 1-shot drops from 66.48% to 65.66% without calibration.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GRACE instead of the strongest listed baseline | few-shot node classification · Cora, CiteSeer, Amazon-Computer, Coauthor-CS, DBLP, CoraFull, and ogbn-arxiv | accuracy (%) | Helped |
| add the high-pass expert (ablation) | GRACE · Cora, CiteSeer, Amazon-Computer, Coauthor-CS, DBLP, CoraFull, and ogbn-arxiv | accuracy (%) | Helped |
| add the low-pass expert (ablation) | GRACE · Cora, CiteSeer, Amazon-Computer, Coauthor-CS, DBLP, CoraFull, and ogbn-arxiv | accuracy (%) | Helped |
| add cross-set distribution calibration (ablation) | GRACE · Cora, CiteSeer, Amazon-Computer, Coauthor-CS, DBLP, CoraFull, and ogbn-arxiv | accuracy (%) | Mixed |
| add adaptive spectrum experts and cross-set distribution calibration (ablation) | GRACE · Cora, CiteSeer, Amazon-Computer, Coauthor-CS, DBLP, CoraFull, and ogbn-arxiv | accuracy (%) | Helped |
Surprising findings
- On Cora 2-way 3-shot, the w/o low-pass variant in the appendix table (78.91%) and w/o cal (82.35%) are close to the full model (82.40%), and on Amazon-Computer 2-way 5-shot the w/o cal variant (94.62%) slightly exceeds the full model in the appendix table (94.66% is the full model, so this is margin
Limitations
- Only low-pass and high-pass experts are considered; broader spectrum experts are not explored.
- Several hyperparameters (bandwidth, temperature, scaling) require tuning.
- Theoretical bound relies on assumed gating error and Wasserstein shift terms that are not empirically verified.
Experimental setup
Datasets: Cora CiteSeer Amazon-Computer Coauthor-CS DBLP CoraFull ogbn-arxiv
Benchmarks: Cora CiteSeer Amazon-Computer Coauthor-CS DBLP CoraFull ogbn-arxiv
Metrics: accuracy standard deviation 95% confidence interval
Baselines: DeepWalk node2vec GCN SGC ProtoNet MAML Meta-GNN GPN G-Meta TENT Meta-GPS X-FNC
Builds on: Prototypical Network GCN MAML GPN
Closest papers by meaning
- HoloNets: Spectral Convolutions do extend to Directed Graphs (ICLR 2024 · similarity 0.88)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.87)
- Unlocking the Potential of Black-box Pre-trained GNNs for Graph Few-shot Learning (AAAI 2025 · similarity 0.87)
- LOHA: Direct Graph Spectral Contrastive Learning Between Low-Pass and High-Pass Views (AAAI 2025 · similarity 0.86)
- SpeAr: A Spectral Approach for Zero-Shot Node Classification (NeurIPS 2024 · similarity 0.86)
- Fast Graph Sharpness-Aware Minimization for Enhancing and Accelerating Few-Shot Node Classification (NeurIPS 2024 · similarity 0.86)
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.