Research map / Graph learning / GNN architectures
GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning
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 GDeR, a dynamic soft data-pruning method for graph learning that uses trainable prototypes on a hyperspherical embedding to sample representative, balanced, and noise-robust graph subsets each epoch.
Problem
Data pruning for graph neural network training speeds up training but degrades performance under class imbalance and noisy/biased training data; no graph-level data pruning method addresses efficiency, balance and robustness together.
Approach
Projects graph embeddings onto a hypersphere modeled by class-specific vMF prototypes with compactness and separation losses, then samples each epoch's training subset using a score combining familiarity, prototype-based Mahalanobis outlier risk, and cluster-minority balancing; unselected samples are sampled with the previous epoch's distribution.
Key results
- On ogbg-molhiv with GraphGPS, GDeR at 50% remaining data reaches 78.7 ROC-AUC, matching the full-dataset 78.7, and at 20% remaining the drop is only 2.2 points versus full data.
- On DHFR imbalanced (1:9) setting, pruning 50% of data yields a 4.3% F1-macro improvement over no-pruning GCN, per the paper.
- Pre-training GraphMAE on ZINC15 with 30% of data gives a 2.81x training speedup, with BBBP improving from 72.04 to 73.57.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GDeR instead of InfoBatch for dynamic pruning | PNA · ogbg-molhiv, 50% remaining | ROC-AUC: 76.3 -> 76.4 | Helped |
| Use GDeR instead of InfoBatch for dynamic pruning | PNA · ogbg-molpcba, 70% remaining | AP: 28.3 -> 28.5 | Helped |
| Use GDeR instead of full-dataset training | GraphGPS · ogbg-molhiv, 50% remaining | ROC-AUC: 78.7 -> 78.7 | No clear effect |
| Use GDeR instead of full-dataset training | GraphGPS · ogbg-molpcba, 50% remaining | AP: 27.2 -> 27.4 | Helped |
| Use GDeR for graph pretraining with 30% of ZINC15 data | GraphMAE · ZINC15 pretraining, training speedup | training speedup: 2.81× | Helped |
| Add familiarity score ω_e to sampling (ablation) | GDeR · Normal, imbalance, and biased scenarios | performance: 79.77_{\pm 2.97} -> 84.21_{\pm 3.40}; 73.78_{\pm 2.96} -> 76.32_{\pm 4.70}; 75.60_{\pm 3.55} -> 77.84_{\pm 2.70} | Helped |
| Add outlier-risk score ω_r to sampling (ablation) | GDeR · Biased scenario | performance: 77.96_{\pm 3.18} vs 77.84_{\pm 2.70} | No clear effect |
| Add balancing score ω_b to sampling (ablation) | GDeR · Imbalance scenario | performance: 73.12_{\pm 2.50} -> 76.32_{\pm 4.70} | Helped |
| Set prototype count K to 1 instead of K=2 (ablation) | GDeR · ogbg-molhiv, GraphGPS, 20% pruning | ROC-AUC: 76.5_{\pm 1.4} -> 75.8_{\pm 1.5} | Hurt |
| Set prototype count K to 4 instead of K=2 (ablation) | GDeR · ogbg-molhiv, GraphGPS, 50% pruning | ROC-AUC: 78.7_{\pm 1.3} -> 78.9_{\pm 0.23} | Helped |
Surprising findings
- InfoBatch, a strong dynamic pruning baseline, degrades by 5.7%-6.4% in biased/noisy training scenarios, while the proposed method does not.
- The paper reports that pruned training on some datasets (e.g., ogbg-molhiv at 30-50% remaining) matches or slightly exceeds full-dataset performance.
Limitations
- Applicability to computer vision and other domains is unexplored.
- Regression and pre-training require virtual classes or ProtNCELoss-based extensions, which are less direct.
- Evaluation is limited to five graph datasets and three GNN backbones.
Experimental setup
Datasets: MUTAG DHFR ogbg-molhiv ogbg-molpcba ZINC15
Benchmarks: MUTAG DHFR ogbg-molhiv ogbg-molpcba MoleculeNet BBBP MoleculeNet ToxCast MoleculeNet BACE
Metrics: ROC-AUC Average Precision Accuracy F1-macro per-epoch time / training speedup
Baselines: InfoBatch Soft Random epsilon-greedy UCB Craig Glister EL2N GraNd Forgetting Herding K-Center DeepFool
Builds on: InfoBatch GraphMAE ProtNCELoss
Closest papers by meaning
- Distilling the Knowledge in Data Pruning (ICML 2025 · similarity 0.90)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.89)
- $\mathbb{D}^2$ Pruning: Message Passing for Balancing Diversity & Difficulty in Data Pruning (ICLR 2024 · similarity 0.89)
- Fast Track to Winning Tickets: Repowering One-Shot Pruning for Graph Neural Networks (AAAI 2025 · similarity 0.89)
- Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty (ICML 2025 · similarity 0.88)
- Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization (NeurIPS 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.