Research map / Graph learning / Graph self-supervised learning
SpeAr: A Spectral Approach for Zero-Shot Node Classification
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 SpeAr, a spectral contrastive loss approximating spectral decomposition of a label- and adjacency-reweighted graph, combined with learnable class prototypes initialized from class semantic vectors and iteratively refined by pseudo-labels, for zero-shot and generalized zero-shot node classification.
Problem
Zero-shot node classification suffers from prediction bias, where unseen-class nodes are misclassified as seen classes; existing methods rely mainly on external semantic knowledge and underuse cluster structure in unlabeled nodes.
Approach
Derive a spectral contrastive loss whose minimization approximates spectral decomposition of the normalized adjacency (reweighted by labels and graph structure), and refine class prototypes iteratively from confident pseudo-labels, initialized with text-based class semantic vectors, in a two-stage training scheme.
Key results
- On Class Split I, SpeAr reaches 60.48% on Cora, 59.72% on Citeseer and 54.22% on C-M10M, versus 48.43%, 40.97% and 45.45% for the best prior baselines.
- On generalized ZNC (Class Split III), SpeAr attains H of 29.98 on Cora and 26.45 on Citeseer, while DGPN gets 0 unseen accuracy and DBiGCN gets H of 2.50 and 8.50.
- On ogbn-arxiv with subgraph extraction, SpeAr gets 30.45% (Split I) and 32.20% (Split II), above GraphCEN's 23.96% and 28.36%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add spectral contrastive loss Lscl (ablation) | SpeAr · Citeseer, Class Split I | accuracy | Helped |
| Use two-stage training instead of omitting stage one (ablation) | SpeAr · Cora, Citeseer, and C-M10M, Class Split I | accuracy | Helped |
| Use text-based rather than label-based class semantic vectors (ablation) | SpeAr · Cora, Citeseer, and C-M10M, Class Split I | accuracy: 49.52 -> 60.48; 48.88 -> 59.72; 47.05 -> 54.22 | Helped |
| Increase the unlabeled-sample loss weight β above the labeled-sample weight α (ablation) | SpeAr · Cora, Citeseer, and C-M10M | accuracy | Helped |
| Vary prototype-update confidence threshold q (ablation) | SpeAr · Cora, Citeseer, and C-M10M | accuracy | Mixed |
| Use SpeAr instead of GraphCEN | zero-shot node classification · Cora, Class Split I | accuracy: 48.43 -> 60.48 | Helped |
| Use SpeAr instead of DBiGCN | zero-shot node classification · Citeseer, Class Split I | accuracy: 40.97 -> 59.72 | Helped |
| Use SpeAr instead of DBiGCN | zero-shot node classification · C-M10M, Class Split I | accuracy: 45.45 -> 54.22 | Helped |
| Use SpeAr instead of GraphCEN | zero-shot node classification · Cora, Class Split II | accuracy: 50.61 -> 58.20 | Helped |
| Use SpeAr instead of DGPN | zero-shot node classification · Citeseer, Class Split II | accuracy: 61.90 -> 75.13 | Helped |
| Use SpeAr instead of DBiGCN | zero-shot node classification · C-M10M, Class Split II | accuracy: 71.86 -> 79.00 | Helped |
| Use SpeAr instead of DBiGCN | generalized zero-shot node classification · Cora, Class Split III | H: 2.50 -> 29.98 | Helped |
Surprising findings
- On GZNC, SpeAr's seen-class accuracy drops sharply (e.g., 34.84% on Cora vs 94.26% for DGPN), trading seen accuracy for unseen accuracy.
- Performance with label-based class semantic vectors is lower than with text-based vectors, with declines of up to 18% for SpeAr.
Limitations
- High computational complexity of pairwise spectral contrastive loss limits large-scale use; mitigated by subgraph training.
- Class Split II tunes parameters on the validation set with unseen classes, which may not reflect strict zero-shot conditions.
- GZNC performance is weak, with low unseen accuracy and large seen-class drop.
Experimental setup
Datasets: Cora Citeseer C-M10M ogbn-arxiv
Benchmarks: Cora Citeseer C-M10M ogbn-arxiv
Metrics: classification accuracy harmonic mean H (generalized ZNC) per-class recall
Baselines: DAP ESZSL ZS-GCN WDVSc Hyperbolic-ZSL DGPN DBiGCN GraphCEN
Builds on: DGPN GraphCEN DBiGCN Eckart-Young-Mirsky theorem spectral contrastive loss
Closest papers by meaning
- FEW-SHOT LEARNING ON GRAPHS VIA SUPER-CLASSES BASED ON GRAPH SPECTRAL MEASURES (ICLR 2020 · similarity 0.88)
- Fast Graph Sharpness-Aware Minimization for Enhancing and Accelerating Few-Shot Node Classification (NeurIPS 2024 · similarity 0.87)
- Graph Few-Shot Learning via Adaptive Spectrum Experts and Cross-Set Distribution Calibration (NeurIPS 2025 · similarity 0.86)
- HoloNets: Spectral Convolutions do extend to Directed Graphs (ICLR 2024 · similarity 0.85)
- Zero-Shot Generalization of GNNs over Distinct Attribute Domains (ICML 2025 · similarity 0.85)
- Revisiting Self-Supervised Heterogeneous Graph Learning from Spectral Clustering Perspective (NeurIPS 2024 · similarity 0.84)
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.