Research map / Graph learning / GNN architectures
On Which Nodes Does GCN Fail? Enhancing GCN From the Node Perspective
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.
Identifies "out-of-control" (OOC) nodes where GCN feature smoothing conflicts with label smoothness, and proposes DaGCN, which adds CVAE-generated virtual neighbors, a dual-similarity KNN graph, and adaptive node-level view fusion with an entropy-reduction loss.
Problem
GCN relies on label smoothness, but a substantial subset of unlabeled nodes (OOC nodes) have feature-smoothed predictions that conflict with label propagation outputs, leading to poor representations, particularly for nodes with few neighbors or far from labeled nodes.
Approach
Locate OOC nodes via disagreement between feature-smoothing predictions and LPA outputs, then enrich them with CVAE-generated virtual neighbors and a DSKNN graph built from both original and virtual features, fused by node-level adaptive weighting with an entropy-reduction regularizer.
Key results
- DaGCN reaches 84.8% on Cora, 75.3% on Citeseer and 81.7% on Pubmed, outperforming GCN by 3.3, 4.4 and 2.7 points.
- DaGCN improves accuracy on OOC nodes by an average of 5.9% over GCN, while UC-node accuracy is similar across methods.
- Under 25% Metattack perturbation on Cora, DaGCN improves over GCN by more than 20%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DaGCN instead of DAGNN | Node classification · Cora | accuracy: 84.4±0.57 -> 84.8± 0.53 | Helped |
| Use DaGCN instead of wGCN | Node classification · Citeseer | accuracy: 73.9 ±0.46 -> 75.3 ± 0.41 | Helped |
| Use DaGCN instead of wGCN | Node classification · Pubmed | accuracy: 80.8±0.25 -> 81.7±0.88 | Helped |
| Use DaGCN instead of wGCN | Node classification · Amazon Computers | accuracy: 83.6±0.86 -> 84.0±1.25 | Helped |
| Use DaGCN instead of wGCN | Node classification · Amazon Photo | accuracy: 92.4±0.18 -> 92.9±0.56 | Helped |
| Use DaGCN instead of DAGNN | Node classification · Coauthor Physics | accuracy: 94.0±0.62 -> 94.3±0.25 | Helped |
| Use DaGCN instead of APPNP | Node classification · Coauthor CS | accuracy: 92.5±0.32 -> 93.4±0.18 | Helped |
Surprising findings
- Current advanced GCNs mainly improve OOC nodes; on UC nodes almost all methods perform similarly to vanilla GCN.
Limitations
- Entropy-reduction term uses a selection function I(.) that must be chosen per dataset.
- Requires many hyperparameters (tau, beta, k, layers, dropout) tuned per dataset.
- Dedicated components for OOC nodes are applied uniformly to all nodes for implementation simplicity.
Experimental setup
Datasets: Cora Citeseer Pubmed Coauthor CS Coauthor Physics Amazon Computers Amazon Photo
Benchmarks: Cora Citeseer Pubmed Coauthor CS Coauthor Physics Amazon Computers Amazon Photo
Metrics: classification accuracy accuracy on UC vs OOC nodes test accuracy under Metattack perturbation generalization gap
Baselines: GCN GAT APPNP GCN-LPA DAGNN AERO-GNN wGNN GCN-Jaccard GCN-SVD RGCN
Builds on: GCN label propagation algorithm APPNP GCNII conditional variational auto-encoder
Closest papers by meaning
- Node Dependent Local Smoothing for Scalable Graph Learning (NeurIPS 2021 · similarity 0.90)
- Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · similarity 0.90)
- On Provable Benefits of Depth in Training Graph Convolutional Networks (NeurIPS 2021 · similarity 0.90)
- Enhancing the Influence of Labels on Unlabeled Nodes in Graph Convolutional Networks (ICML 2025 · similarity 0.89)
- L2-GCN: Layer-Wise and Learned Efficient Training of Graph Convolutional Networks (CVPR 2020 · similarity 0.89)
- GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights (NeurIPS 2025 · similarity 0.87)
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.