Research map / Graph learning / GNN architectures
Graph structure learning
Learn or refine graph structure to improve node classification.
201 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in GNN architectures, keyword cluster “gnn · explanations · trained”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 6 tracked benchmarks.
61 papers in 2023–24 and 75 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 12.9% to 14.8%.
Open questions the data points to
- Cora (accuracy): best went from 89.89 (up to 2024) to 91.2 in 2025 — Tokenphormer: Structure-aware Multi-token Graph Transformer for Node Classification (AAAI 2025 · 6 citations)
- Papers report IDGL on Cora (accuracy (%)) anywhere from 70.9 to 83.23: comparisons against it depend on whose number you take.
What looks settled
- Citeseer (accuracy): best up to 2024 83.51, best in 2025–26 81.27 from 6 papers, at most 0.2% better, if at all.
- Pubmed (accuracy): best up to 2024 89.79, best in 2025–26 90.64 from 5 papers, less than 1% better.
No longer reported by papers here: IMDB-B (accuracy, last in 2024); IMDB-M (accuracy, last in 2024).
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Deep Collective Inference (AAAI 2017 · 29 citations)
- 2018 · Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · 2,635 citations)
- 2019 · Explainability Methods for Graph Convolutional Neural Networks (CVPR 2019 · 531 citations)
- 2020 · ASAP: Adaptive Structure Aware Pooling for Learning Hierarchical Graph Representations (AAAI 2020 · 306 citations)
- 2021 · Data Augmentation for Graph Neural Networks (AAAI 2021 · 350 citations)
- 2022 · Graph Structure Learning with Variational Information Bottleneck (AAAI 2022 · 162 citations)
- 2023 · Federated Learning on Non-IID Graphs via Structural Knowledge Sharing (AAAI 2023 · 165 citations)
- 2024 · HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · 36 citations)
- 2025 · When Hypergraph Meets Heterophily: New Benchmark Datasets and Baseline (AAAI 2025 · 34 citations)
- 2026 · Hierarchical Frequency-Decomposition Graph Neural Networks for Road Network Representation Learning (AAAI 2026 · 4 citations)
Earliest papers here
- Deep Collective Inference (AAAI 2017 · 29 citations)
- Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · 2,635 citations)
- An End-to-End Deep Learning Architecture for Graph Classification (AAAI 2018 · 1,505 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
Cora · accuracy (higher is better)
25 papers, 2019–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 85.5 | Semi-Supervised Learning With Graph Learning-Convolutional Networks (CVPR 2019 · 484 citations) |
| 2021 | 89.89 | AutoGEL: An Automated Graph Neural Network with Explicit Link Information (NeurIPS 2021 · 20 citations) |
| 2025 | 91.2 | Tokenphormer: Structure-aware Multi-token Graph Transformer for Node Classification (AAAI 2025 · 6 citations) |
Citeseer · accuracy (higher is better)
24 papers, 2019–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 73.6 | GMNN: Graph Markov Neural Networks (ICML 2019 · 160 citations) |
| 2020 | 75.16 | When Does Self-Supervision Help Graph Convolutional Networks? (ICML 2020 · 95 citations) |
| 2023 | 79.79 | Curriculum Learning for Graph Neural Networks: Which Edges Should We Learn First (NeurIPS 2023 · 3 citations) |
| 2024 | 83.51 | Rethinking Independent Cross-Entropy Loss For Graph-Structured Data (ICML 2024 · 1 citation) |
Pubmed · accuracy (higher is better)
23 papers, 2019–2025 · small gains only
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 81.9 | GMNN: Graph Markov Neural Networks (ICML 2019 · 160 citations) |
| 2021 | 89.68 | AutoGEL: An Automated Graph Neural Network with Explicit Link Information (NeurIPS 2021 · 20 citations) |
| 2023 | 89.79 | Curriculum Learning for Graph Neural Networks: Which Edges Should We Learn First (NeurIPS 2023 · 3 citations) |
| 2025 | 90.64 | N2GON: Neural Networks for Graph-of-Net with Position Awareness (ICML 2025) |
PROTEINS · accuracy (higher is better)
13 papers, 2018–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 76.25 | Hierarchical Graph Representation Learning with Differentiable Pooling (NeurIPS 2018 · 882 citations) |
| 2020 | 79.91 | Graph Cross Networks with Vertex Infomax Pooling (NeurIPS 2020 · 4 citations) |
| 2021 | 82.68 | AutoGEL: An Automated Graph Neural Network with Explicit Link Information (NeurIPS 2021 · 20 citations) |
IMDB-B · accuracy (higher is better)
11 papers, 2018–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 70.03 | An End-to-End Deep Learning Architecture for Graph Classification (AAAI 2018 · 1,505 citations) |
| 2020 | 78.6 | Graph Cross Networks with Vertex Infomax Pooling (NeurIPS 2020 · 4 citations) |
| 2021 | 81.2 | AutoGEL: An Automated Graph Neural Network with Explicit Link Information (NeurIPS 2021 · 20 citations) |
IMDB-M · accuracy (higher is better)
11 papers, 2018–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 47.83 | An End-to-End Deep Learning Architecture for Graph Classification (AAAI 2018 · 1,505 citations) |
| 2020 | 55.2 | Graph Cross Networks with Vertex Infomax Pooling (NeurIPS 2020 · 4 citations) |
| 2021 | 56.8 | AutoGEL: An Automated Graph Neural Network with Explicit Link Information (NeurIPS 2021 · 20 citations) |
What has been tried
155 papers here report 1,652 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 7%, hurt 18%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add a graph template | Tab. 3 benchmark settings · performance | Helped | HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · 36 citations) | |
| Use labeled-unlabeled rather than labeled-labeled pairing | Cora · accuracy | 81.85 -> 83.64 | Helped | NodeMixup: Tackling Under-Reaching for Graph Neural Networks (AAAI 2024 · 25 citations) |
| Add feature-level contrastive learning | TOX-171 · ACC | 47.95 -> 51.46 | Helped | Deep Contrastive Graph Learning with Clustering-Oriented Guidance (AAAI 2024 · 19 citations) |
| Add graph structure information | MolHIV · AUC | 0.728 -> 0.790 | Helped | G-Adapter: Towards Structure-Aware Parameter-Efficient Transfer Learning for Graph Transformer Networks (AAAI 2024 · 15 citations) |
| Remove residual connections by setting α to 0 | Node classification · accuracy | Hurt | ASWT-SGNN: Adaptive Spectral Wavelet Transform-Based Self-Supervised Graph Neural Network (AAAI 2024 · 13 citations) | |
| Add graphon reconstruction loss to GIN without pre-training | PROTEINS and ENZYMES (original node features) · Accuracy | 73.3±4.0 -> 74.9±3.1; 59.6±4.5 -> 66.0±6.2 | Helped | Fine-Tuning Graph Neural Networks by Preserving Graph Generative Patterns (AAAI 2024 · 13 citations) |
| add hierarchical subclass contrastive learning | Synthie, ENZYMES, MNIST, Letter-high, Letter-low, COIL-DEL · accuracy (%) | Helped | Cluster-guided Contrastive Class-imbalanced Graph Classification (AAAI 2025 · 11 citations) | |
| add common-neighbor pairwise features to GAE | OGB datasets · link prediction score | 70% | Helped | Neural Common Neighbor with Completion for Link Prediction (ICLR 2024 · 11 citations) |
| add domain adaptation | BBBP · ROC-AUC | 86.71±0.74 -> 88.75±0.49 | Helped | Pre-Training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2025 · 10 citations) |
| add graph token embeddings to provide graph information to the LLM | Cross-dataset and cross-task evaluations · performance | Helped | LLMs as Zero-shot Graph Learners: Alignment of GNN Representations with LLM Token Embeddings (NeurIPS 2024 · 10 citations) | |
| add the line-graph branch to the supervised hypergraph branch | PROTEINS · accuracy (%) | 69.1±1.0 -> 70.1±0.9 | Helped | Hypergraph-enhanced Dual Semi-supervised Graph Classification (ICML 2024 · 8 citations) |
| Add centrality encoding | Bitcoin-OTC and Bitcoin-Alpha · accuracy | Helped | Self-Explainable Graph Transformer for Link Sign Prediction (AAAI 2025 · 7 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| GCN | PubMed | accuracy | 78.98 – 83.51 | 3 papers, e.g. NeurIPS 2023 and AAAI 2023 |
| IDGL | Cora | accuracy (%) | 70.9 – 83.23 | 2 papers, e.g. NeurIPS 2021 and AAAI 2023 |
| DGI | Pubmed | accuracy | 76.8 – 84.29 | 2 papers, e.g. ICML 2019 and ICLR 2022 |
| SAGPool | PROTEINS | accuracy (%) | 71.56 – 78.28 | 2 papers, e.g. NeurIPS 2021 and NeurIPS 2020 |
| GIN | MUTAG | accuracy | 81.39 – 87.19 | 2 papers, e.g. ICLR 2021 and NeurIPS 2023 |
What reviewers will expect
Datasets and baselines named in the 44 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Cora · 15 of 44 papers
- Citeseer · 14 of 44 papers
- Pubmed · 10 of 44 papers
- DBLP · 6 of 44 papers
- ogbn-arxiv · 5 of 44 papers
- ENZYMES · 4 of 44 papers
- ACM · 4 of 44 papers
- PROTEINS · 4 of 44 papers
Baselines
- GCN · 14 of 44 papers
- GAT · 12 of 44 papers
- GraphSAGE · 6 of 44 papers
- HAN · 4 of 44 papers
- GraphPrompt · 4 of 44 papers
- GraphCL · 3 of 44 papers
- HGT · 3 of 44 papers
- MAGNN · 3 of 44 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 136 papers; from OpenAlex and the papers' affiliations).
- Zhejiang University · 7 papers
- Peking University · 7 papers
- Beihang University · 6 papers
- Beijing University of Posts and Telecommunications · 5 papers
- Chinese Academy of Sciences · 5 papers
- Amazon · 5 papers
- The Hong Kong Polytechnic University · 4 papers
- National University of Singapore · 4 papers
- University of Electronic Science and Technology of China · 4 papers
- Michigan State University · 4 papers
Use it
Download BibTeX of 40 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- GNN generalization theory (187 papers)
- Simple and scalable GNNs (131 papers)
- Robustness, privacy, and fairness of GNNs (94 papers)
- Physics-inspired GNNs (81 papers)
- GNN explainability (64 papers)
Elsewhere in the topic
- Expressive equivariant message passing (109 papers)
- Theory of graph convolutional networks (83 papers)
- Deep graph convolution and simplification (82 papers)
- Graph rewiring and scalable message passing (72 papers)
- Weisfeiler-Leman expressivity hierarchies (66 papers)
- Structural GNN variants (65 papers)
- Oversmoothing and heterophily (60 papers)
- Substructure-aware GNN expressivity (50 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).