Research map / Graph learning / Graph self-supervised learning
GC-Bench: An Open and Unified Benchmark for Graph Condensation
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 GC-Bench, a unified benchmark and open-source library evaluating 12 graph condensation methods on 12 node- and graph-level datasets for effectiveness, transferability, and efficiency.
Problem
Graph condensation methods have proliferated, but there is no comprehensive, unified evaluation of their effectiveness, transferability, and efficiency across node-level and graph-level tasks.
Approach
A unified benchmark that re-runs 12 GC methods under common settings across a broader range of condensation ratios, 7 node-level and 5 graph-level datasets, and analyzes structure, cross-task and cross-backbone transfer, initialization, and time/memory costs.
Key results
- Node-level GC methods (e.g., GEOM, SFGC) reach near-lossless accuracy in 5 of 7 node-level cases, e.g., Citeseer at r=0.18% and ACM at r=0.003%.
- On graph-level tasks a large condensation ratio does not necessarily improve accuracy; KiDD with GIN wins in 18 of 25 graph-level cases.
- Condensed datasets transfer poorly across tasks (link prediction, node clustering, anomaly detection), and graph transformers as condensation backbones yield suboptimal results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use gradient-matching or trajectory-matching condensation rather than distribution matching | Node-level graph condensation · 7 node-classification datasets | accuracy | Helped |
| Increase the condensation ratio beyond the method's effective range | Deep graph-condensation methods · Node- and graph-level tasks | performance | Hurt |
| Transfer node-classification-condensed graphs to other downstream tasks | Graph condensation methods · Citeseer link prediction, node clustering, and anomaly detection | performance | Hurt |
| Use Graph Transformer rather than SGC as the condensation backbone | Graph condensation methods · Cora, 2.6% condensation ratio; cross-architecture evaluation | accuracy | Hurt |
| Increase the number of gradient-matching steps | GCond · Cora, 2.6% condensation ratio; cross-architecture evaluation | accuracy | Hurt |
| Initialize GCond with Center rather than Random Sample (ablation) | GCond · Cora | time to reach the same accuracy: approximately 25% | Helped |
| Initialize GCond with Center rather than Random Noise (ablation) | GCond · Cora | time to reach the same accuracy: 71% | Helped |
| Change the initialization strategy (ablation) | GCond · Cora and arxiv | final accuracy | No clear effect |
Surprising findings
- A larger condensation ratio does not necessarily yield better performance for current GC methods.
- Graph transformers, despite strong general performance, perform suboptimally when used for graph condensation.
Limitations
- Evaluation restricted to node classification and graph classification for the main comparisons, and only 12 datasets.
- Some methods (e.g., Mirage, deep GC methods) cannot hit exact ratios or OOM on large datasets, leaving missing entries.
Experimental setup
Datasets: Cora Citeseer ogbn-arxiv Flickr Reddit ACM DBLP NCI1 DD ogbg-molbace ogbg-molbbbp ogbg-molhiv
Benchmarks: Cora Citeseer ogbn-arxiv Flickr Reddit ACM DBLP NCI1 DD ogbg-molbace ogbg-molbbbp ogbg-molhiv
Metrics: accuracy ROC-AUC training time peak CPU memory peak GPU memory
Baselines: Random Herding K-Center GCond DosCond SGDD GCDM DM SFGC GEOM KiDD Mirage
Builds on: GCond DosCond SGDD SFGC GEOM KiDD Mirage
Closest papers by meaning
- GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights (NeurIPS 2025 · similarity 0.93)
- Bonsai: Gradient-free Graph Condensation for Node Classification (ICLR 2025 · similarity 0.90)
- Adapting Precomputed Features for Efficient Graph Condensation (ICML 2025 · similarity 0.89)
- Navigating Complexity: Toward Lossless Graph Condensation via Expanding Window Matching (ICML 2024 · similarity 0.88)
- Towards Pre-trained Graph Condensation via Optimal Transport (NeurIPS 2025 · similarity 0.88)
- Bi-Directional Multi-Scale Graph Dataset Condensation via Information Bottleneck (AAAI 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.