Research map / Graph learning / GNN architectures
Adapting Precomputed Features for Efficient 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.
Proposes GCPA, a structure-free graph condensation method that replaces costly trajectory matching with one-time GNN-style feature precomputation plus a diversity-aware contrastive feature adaptation stage, achieving comparable accuracy with 96x-2,455x speedups.
Problem
Trajectory-matching graph condensation methods achieve SOTA accuracy but require hundreds of repeated GNN retraining rounds for trajectory collection, making condensation prohibitively slow on large graphs.
Approach
Bypass trajectory matching entirely: precompute structure-aware features via K-hop diffusion, condense them by class-wise uniform sampling and averaging, then adapt the synthetic features with a contrastive class-alignment loss plus an intra-class diversity term, keeping the precomputed features as a fixed residual anchor.
Key results
- The precomputation-only variant matches or surpasses 5 of 9 baselines in Table 2 while taking under 60 seconds.
- Full GCPA achieves SOTA on 4 of 7 datasets and is 96x to 2,455x faster than GEOM (e.g., Products 104s vs 1,912,105s total condensation time for GEOM at r=0.05%: Ours 2,985s vs GEOM 1,912,105s).
- Cross-architecture results: GCPA attains the best average accuracy on Cora (78.3) and Reddit (74.1) across seven evaluation architectures.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add structure-based precomputation (ablation) | GCPA · CiteSeer, Cora, PubMed, Arxiv, Products, Flickr, Reddit | accuracy: CiteSeer: 69.3±0.6 -> 74.8±0.3; Cora: 74.4±0.4 -> 82.9±1.0; PubMed: 77.9±0.3 -> 80.9±0.3; Arxiv: 64.2±0.6 -> 67.7±0.2; P | Helped |
| Add semantic-based precomputation (ablation) | GCPA · CiteSeer, Cora, PubMed, Arxiv, Products, Flickr, Reddit | accuracy: CiteSeer: 67.8±0.2 -> 74.8±0.3; Cora: 79.1±0.7 -> 82.9±1.0; PubMed: 78.8±0.4 -> 80.9±0.3; Arxiv: 67.3±0.3 -> 67.7±0.2; P | Helped |
| Add both structure-based and semantic-based precomputation (ablation) | GCPA · CiteSeer, Cora, PubMed, Arxiv, Products, Flickr, Reddit | accuracy: CiteSeer: 62.1±0.3 -> 74.8±0.3; Cora: 72.2±0.5 -> 82.9±1.0; PubMed: 76.1±0.5 -> 80.9±0.3; Arxiv: 63.9±0.2 -> 67.7±0.2; P | Helped |
| Increase diversity constraint weight from γ=0 to γ=0.1 (ablation) | GCPA · CiteSeer, Cora, PubMed, Arxiv, Products, Flickr, Reddit | accuracy: CiteSeer: 72.1±0.1 -> 74.8±0.3; Cora: 81.4±0.2 -> 80.3±0.6; PubMed: 79.9±0.5 -> 80.9±0.3; Arxiv: 64.6±0.4 -> 67.7±0.2; P | Mixed |
| Set diversity constraint weight to γ=1 (ablation) | GCPA · CiteSeer, Cora, PubMed, Arxiv, Products, Flickr, Reddit | accuracy: CiteSeer: 74.8±0.3 -> 74.2±0.2; Cora: 82.9±1.0 -> 79.7±0.7; PubMed: 80.9±0.3 -> 79.3±0.4; Arxiv: 67.7±0.2 -> 66.1±0.5; P | Mixed |
| Use the adaptation stage after precomputation | GCPA · Ogbn-arxiv, Flickr | accuracy | Helped |
Surprising findings
- Precomputation alone, with no learning, already matches or beats 5 of 9 baselines on node classification.
- Removing structure-based aggregation typically hurts more than removing semantic aggregation, especially on transductive datasets.
Limitations
- Gains depend on the precomputation design; the authors note only simple diffusion/averaging was explored.
- Performance is mixed on some datasets (e.g., Flickr and Products are below GEOM or SGDD in some ratios), and efficiency comparisons rely on reported or re-run timings.
Experimental setup
Datasets: CiteSeer Cora PubMed Ogbn-arxiv Ogbn-products Flickr Reddit
Benchmarks: CiteSeer Cora PubMed Ogbn-arxiv Ogbn-products Flickr Reddit
Metrics: node classification accuracy condensation time (s)
Baselines: K-Center GCond SGDD GCDM SimGC EXGC CGC SFGC GEOM
Builds on: GCond SFGC GEOM SimGC APPNP-style diffusion
Closest papers by meaning
- Towards Pre-trained Graph Condensation via Optimal Transport (NeurIPS 2025 · similarity 0.92)
- Navigating Complexity: Toward Lossless Graph Condensation via Expanding Window Matching (ICML 2024 · similarity 0.90)
- GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights (NeurIPS 2025 · similarity 0.90)
- Bonsai: Gradient-free Graph Condensation for Node Classification (ICLR 2025 · similarity 0.89)
- GC-Bench: An Open and Unified Benchmark for Graph Condensation (NeurIPS 2024 · similarity 0.89)
- Structure Balance and Gradient Matching-Based Signed Graph Condensation (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.