Research map / Graph learning / GNN architectures
Aggregation Buffer: Revisiting DropEdge with a New Parameter Block
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.
Revisits DropEdge, argues its gains are limited by GNN aggregation's lack of discrepancy bounds, and proposes Aggregation Buffer, a frozen-backbone add-on block trained with DropEdge and a robustness-controlled loss.
Problem
DropEdge gives limited gains for GNN supervised node classification; the paper asks why and how to improve GNN robustness to edge/structural perturbations (degree bias, structural disparity).
Approach
Theoretically shows GNN aggregation lacks a constant discrepancy bound under different adjacency matrices, so DropEdge's robustness term cannot be optimized; adds a degree-normalized linear parameter block (D+I)^{-1} H W to each layer of a frozen pre-trained GNN and trains it with DropEdge plus a robustness-controlled loss (bias via distillation on training nodes, robustness via KL between clean a
Key results
- GCN_B achieves the highest accuracy in 9 and second-best in 3 of 12 datasets overall (e.g., PubMed 87.56 vs GCN 86.48 and best baseline 86.56).
- Aggregation Buffer improves GCN on CS from 91.85 to 93.54 and on Arxiv from 71.80 to 72.43.
- Integrating the buffer into SAGE, GAT, SGC, and GIN improves accuracy on Pubmed, CS, Arxiv, and Chameleon in all reported cases (e.g., GIN on Pubmed 85.42 to 87.18).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Aggregation Buffer to GCN | GCN · Cora, Citeseer, PubMed, Wiki-CS, Photo, Computer, CS, Physics, Arxiv, Actor, Squirrel, Chameleon | overall accuracy (%): 83.44 -> 84.84; 72.45 -> 73.32; 86.48 -> 87.56; 80.26 -> 80.75; 92.21 -> 92.44; 88.24 -> 88.76; 91.85 -> 93.54; 95.18 -> | Helped |
| add Aggregation Buffer to SAGE, GAT, SGC, or GIN | SAGE, GAT, SGC, GIN · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.07 -> 88.09; 92.44 -> 93.36; 70.92 -> 71.16; 37.34 -> 37.85; 85.64 -> 87.47; 90.50 -> 93.09; 71.86 -> 72.26; 38.54 -> | Helped |
| unfreeze the pretrained GCN parameters (ablation) | GCN_B · PubMed, CS, Chameleon | accuracy (%): 87.56 -> 87.50; 93.54 -> 93.50; 40.96 -> 40.77 | Hurt |
| remove Aggregation Buffer (ablation) | GCN_B · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.56 -> 86.82; 93.54 -> 91.67; 72.43 -> 72.21; 40.96 -> 40.35 | Hurt |
| remove pretraining (ablation) | GCN_B · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.56 -> 87.22; 93.54 -> 92.78; 72.43 -> 71.15; 40.96 -> 39.39 | Mixed |
| replace Aggregation Buffer with standard AGG (ablation) | GCN_B · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.56 -> 86.82; 93.54 -> 91.63; 72.43 -> 72.27; 40.96 -> 40.69 | Hurt |
| replace Aggregation Buffer with JKNet (ablation) | GCN_B · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.56 -> 87.45; 93.54 -> 93.36; 72.43 -> 72.19; 40.96 -> 40.29 | Mixed |
| replace Aggregation Buffer with residual connection (ablation) | GCN_B · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.56 -> 87.46; 93.54 -> 92.05; 72.43 -> 72.29; 40.96 -> 39.77 | Mixed |
| restrict the robustness term to training nodes (ablation) | GCN_B training objective · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.56 -> 86.89; 93.54 -> 92.67; 72.43 -> 72.26; 40.96 -> 40.31 | Hurt |
| replace L_RC with pseudo-label loss (ablation) | GCN_B training objective · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.56 -> 86.62; 93.54 -> 92.48; 72.43 -> 71.84; 40.96 -> 40.14 | Hurt |
| replace L_RC with self-distillation loss (ablation) | GCN_B training objective · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.56 -> 86.15; 93.54 -> 92.02; 72.43 -> 72.18; 40.96 -> 40.22 | Hurt |
| replace L_RC with cross-entropy loss (ablation) | GCN_B training objective · PubMed, CS, Arxiv, Chameleon | accuracy (%): 87.56 -> 86.67; 93.54 -> 93.29; 72.43 -> 71.94; 40.96 -> 40.76 | Hurt |
Surprising findings
- Standard aggregation (GCN's AGG) as the add-on block gives no improvement on PubMed and degrades CS relative to base GCN.
- Well-tuned MLPs outperform GNNs on CS and Actor datasets.
Limitations
- Two-step approach relies on a pre-trained GNN, so performance depends on pre-trained knowledge.
- Gains on some datasets are within standard deviation (e.g., Chameleon, Squirrel); buffer hyperparameters are not tuned per dataset.
- Requires tuning of dropout, DropEdge ratio, and lambda.
Experimental setup
Datasets: Cora Citeseer PubMed Wiki-CS Amazon Photo Amazon Computers Coauthor CS Coauthor Physics ogbn-arxiv Actor Squirrel Chameleon
Benchmarks: Cora Citeseer PubMed Wiki-CS Amazon Photo Amazon Computers Coauthor CS Coauthor Physics ogbn-arxiv Actor Squirrel Chameleon
Metrics: accuracy accuracy on head nodes accuracy on tail nodes accuracy on homophilous nodes accuracy on heterophilous nodes
Baselines: GCN MLP DropEdge DropNode DropMessage TUNEUP GraphPatcher
Builds on: DropEdge GCN TUNEUP DropMessage
Closest papers by meaning
- DropEdge: Towards Deep Graph Convolutional Networks on Node Classification (ICLR 2020 · similarity 0.89)
- Effects of Dropout on Performance in Long-range Graph Learning Tasks (NeurIPS 2025 · similarity 0.87)
- Geometry-Aware Edge Pooling for Graph Neural Networks (NeurIPS 2025 · similarity 0.86)
- Integrating Co-Training with Edge Discrimination to Enhance Graph Neural Networks Under Heterophily (AAAI 2025 · similarity 0.86)
- Graph Neural Network Generalization With Gaussian Mixture Model Based Augmentation (ICML 2025 · similarity 0.85)
- Moderate Message Passing Improves Calibration: A Universal Way to Mitigate Confidence Bias in Graph Neural Networks (AAAI 2024 · similarity 0.85)
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.