Research map / Graph learning / GNN architectures
GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights
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.
Introduces GC4NC, a unified benchmark for graph condensation on node classification that evaluates performance, efficiency, privacy (MIA), denoising, NAS, transferability, and design choices across ~20 methods.
Problem
Graph condensation methods for node classification are evaluated inconsistently (different validation models, test-set selection, validation frequency), and key dimensions such as privacy, denoising, NAS effectiveness, and transferability are under-explored.
Approach
A unified evaluation protocol (fixed GCN validator, limited validation frequency, selection by validation accuracy) applied across many coreset, coarsening, and condensation methods, extended to multi-dimensional evaluation and analysis of design choices such as initialization, structure synthesis, and graph property preservation.
Key results
- Trajectory-matching methods GEOM and SFGC give the best condensation accuracy (e.g., 82.82% on Cora at 2.6%), but are costly in preprocessing and disk memory.
- Under structural and adversarial structural noise, GC methods such as GCond outperform GCN trained on the corrupted whole graph (e.g., 65.64% vs 57.63% on Citeseer), but they do not denoise feature noise.
- DosCond, which lacks inner optimization, yields poor NAS correlation (Pearson 0.14 accuracy correlation vs 0.65-0.72 for trajectory-matching methods).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use trajectory matching rather than other condensation objectives | Graph condensation methods · Node classification across benchmark datasets and reduction rates | test accuracy | Helped |
| Use trajectory matching | Graph condensation for NAS · NAS | accuracy and rank correlation: 0.72 accuracy correlation; 0.71 rank correlation | Helped |
| Remove inner optimization (as in DosCond) | Graph condensation for NAS · NAS | accuracy correlation: 0.14 | Hurt |
| Use structure-based rather than structure-free condensation | Graph condensation · Structural and adversarial structural noise | test accuracy under noise | Helped |
| Use graph condensation rather than whole-graph training | GC methods · Structural and adversarial structural noise | test accuracy | Helped |
| Apply graph condensation to feature-corrupted graphs | GC methods · Feature noise | test accuracy | No clear effect |
| Use structure-free instead of structure-based condensation | Graph condensation · Performance and resource use | efficiency | Helped |
| Use structure-free instead of structure-based condensation | Graph condensation · Structural noise | test accuracy under noise | Hurt |
| Increase the reduction rate | Graph condensation · Larger reduction rates | scalability | Hurt |
| Use a different data initialization strategy (ablation) | Graph condensation methods · Node classification across datasets | test accuracy | Mixed |
| Initialize with Averaging (ablation) | Graph condensation methods · Large datasets | test accuracy | Hurt |
| Initialize with KCenter (ablation) | Graph condensation methods · Most datasets | test accuracy | Helped |
Surprising findings
- Averaging, a learning-free coarsening baseline, performs best on Yelp among all methods.
- Graph condensation can keep good performance at a single instance per class, whereas image condensation drops sharply at similar reduction rates.
Limitations
- Focuses only on node classification and on a limited set of datasets (seven) compared to other benchmarks.
- GDEM could only be reproduced on small datasets with the provided eigendecomposition, so its results on large graphs are missing.
- Privacy evaluation uses only MIA; differential privacy and other attacks are not studied.
Experimental setup
Datasets: Cora Citeseer Pubmed Arxiv Flickr Reddit Yelp
Benchmarks: Cora Citeseer Pubmed Arxiv Flickr Reddit Yelp
Metrics: test accuracy F1-macro running time peak and GPU memory disk memory membership inference attack accuracy Pearson correlation for NAS rank correlation
Baselines: Random K-Center Herding Averaging VNG GCDM GCSNTK GDEM SGDD
Builds on: GCond DosCond SFGC GEOM MSGC
Closest papers by meaning
- GC-Bench: An Open and Unified Benchmark for Graph Condensation (NeurIPS 2024 · similarity 0.93)
- Bonsai: Gradient-free Graph Condensation for Node Classification (ICLR 2025 · similarity 0.93)
- Towards Pre-trained Graph Condensation via Optimal Transport (NeurIPS 2025 · similarity 0.90)
- Adapting Precomputed Features for Efficient Graph Condensation (ICML 2025 · similarity 0.90)
- Robust Graph Condensation via Classification Complexity Mitigation (NeurIPS 2025 · similarity 0.90)
- Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 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.