Research map / Graph learning / Graph self-supervised learning
Semi-supervised node representation learning
Learn node embeddings and classifiers on graphs with GCNs and inductive methods.
214 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Graph self-supervised learning, keyword cluster “classification · semi supervised · representation”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Leveling off Several papers from 2025–26 still report on 5 tracked benchmarks, but the best result has barely moved since 2024.
54 papers in 2023–24 and 57 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 15.7% to 9.8%.
What looks settled
- Cora (accuracy): best up to 2024 90.49, best in 2025–26 88.48 from 3 papers, at most 0.2% better, if at all.
- Citeseer (accuracy): best up to 2024 78.53, best in 2025–26 77.53 from 2 papers, at most 0.2% better, if at all.
- Pubmed (accuracy): best up to 2024 89.44, best in 2025–26 89.96 from 2 papers, less than 1% better.
- MUTAG (accuracy): best up to 2024 94.68, best in 2025–26 90.99 from 3 papers, at most 0.2% better, if at all.
- PROTEINS (accuracy): best up to 2024 80.12, best in 2025–26 77.47 from 2 papers, at most 0.2% better, if at all.
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
- 2016 · Deep Neural Networks for Learning Graph Representations (AAAI 2016 · 1,063 citations)
- 2017 · Semi-Supervised Classification with Graph Convolutional Networks (ICLR 2017 · 7,975 citations)
- 2018 · Representation Learning on Graphs with Jumping Knowledge Networks (ICML 2018 · 970 citations)
- 2019 · Hypergraph Neural Networks (AAAI 2019 · 1,751 citations)
- 2020 · Contrastive Multi-View Representation Learning on Graphs (ICML 2020 · 450 citations)
- 2021 · Deep Fusion Clustering Network (AAAI 2021 · 242 citations)
- 2022 · Deep Graph Clustering via Dual Correlation Reduction (AAAI 2022 · 240 citations)
- 2023 · SGFormer: Simplifying and Empowering Transformers for Large-Graph Representations (NeurIPS 2023 · 37 citations)
- 2024 · Gramformer: Learning Crowd Counting via Graph-Modulated Transformer (AAAI 2024 · 45 citations)
- 2025 · Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach (AAAI 2025 · 5 citations)
- 2026 · High-Pass Matters: Theoretical Insights and Sheaflet-Based Design for Hypergraph Neural Networks (AAAI 2026 · 5 citations)
Earliest papers here
- Deep Neural Networks for Learning Graph Representations (AAAI 2016 · 1,063 citations)
- Diffusion-Convolutional Neural Networks (NeurIPS 2016 · 566 citations)
- Revisiting Semi-Supervised Learning with Graph Embeddings (ICML 2016 · 281 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)
24 papers, 2017–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 81.5 | Semi-Supervised Classification with Graph Convolutional Networks (ICLR 2017 · 7,975 citations) |
| 2018 | 89.6 | Representation Learning on Graphs with Jumping Knowledge Networks (ICML 2018 · 970 citations) |
| 2023 | 90.49 | Transformers over Directed Acyclic Graphs (NeurIPS 2023 · 4 citations) |
Citeseer · accuracy (higher is better)
22 papers, 2017–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 70.3 | Semi-Supervised Classification with Graph Convolutional Networks (ICLR 2017 · 7,975 citations) |
| 2018 | 78.3 | Representation Learning on Graphs with Jumping Knowledge Networks (ICML 2018 · 970 citations) |
| 2024 | 78.53 | Refining Latent Homophilic Structures over Heterophilic Graphs for Robust Graph Convolution Networks (AAAI 2024 · 11 citations) |
Pubmed · accuracy (higher is better)
22 papers, 2017–2025 · small gains only
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 79 | Semi-Supervised Classification with Graph Convolutional Networks (ICLR 2017 · 7,975 citations) |
| 2018 | 82.4 | Bayesian Semi-supervised Learning with Graph Gaussian Processes (NeurIPS 2018 · 17 citations) |
| 2020 | 83.1 | Graph inference learning for semi-supervised classification (ICLR 2020 · 18 citations) |
| 2023 | 89.44 | Transformers over Directed Acyclic Graphs (NeurIPS 2023 · 4 citations) |
| 2025 | 89.96 | Unifying and Enhancing Graph Transformers via a Hierarchical Mask Framework (NeurIPS 2025 · 0 citations) |
MUTAG · accuracy (higher is better)
15 papers, 2017–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 89.44 | Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on Graphs (CVPR 2017 · 1,334 citations) |
| 2019 | 90.51 | A Persistent Weisfeiler-Lehman Procedure for Graph Classification (ICML 2019 · 41 citations) |
| 2021 | 94.68 | Deep Wasserstein Graph Discriminant Learning for Graph Classification (AAAI 2021 · 10 citations) |
PROTEINS · accuracy (higher is better)
12 papers, 2019–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 75.86 | A Persistent Weisfeiler-Lehman Procedure for Graph Classification (ICML 2019 · 41 citations) |
| 2021 | 77.29 | Deep Wasserstein Graph Discriminant Learning for Graph Classification (AAAI 2021 · 10 citations) |
| 2024 | 80.12 | Transitivity-Preserving Graph Representation Learning for Bridging Local Connectivity and Role-Based Similarity (AAAI 2024 · 12 citations) |
NCI1 · accuracy (higher is better)
10 papers, 2017–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 83.8 | Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on Graphs (CVPR 2017 · 1,334 citations) |
| 2019 | 85.62 | A Persistent Weisfeiler-Lehman Procedure for Graph Classification (ICML 2019 · 41 citations) |
What has been tried
171 papers here report 1,854 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 9%, hurt 17%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add EWR to the VGG+IAL baseline | UCF-QNRF · MAE | 84.7 -> 79.3 | Helped | Gramformer: Learning Crowd Counting via Graph-Modulated Transformer (AAAI 2024 · 45 citations) |
| Add graph contrastive loss for local graph learning | Cora {gr, lr}={.1, .3} · accuracy | 53.93±0.66 -> 81.73±1.42 | Helped | Robust Node Classification on Graph Data with Graph and Label Noise (AAAI 2024 · 32 citations) |
| Add hypergraph reconstruction loss | Hypervertex classification: Citeseer, Pubmed, Cora, DBLP · accuracy | 70.3 ± 0.9 -> 73.1 ± 0.7; 80.0 ± 2.2 -> 81.0 ± 1.5; 65.7 ± 0.6 -> 67.1 ± 0.6; 88.4 ± 0.2 -> 89.1 ± 0.1 | Helped | Hypergraph Joint Representation Learning for Hypervertices and Hyperedges via Cross Expansion (AAAI 2024 · 18 citations) |
| Remove the causal-information objective max I(C; Y) | MUTAG, PROTEINS, Graph-SST2 · classification performance | Hurt | A Twist for Graph Classification: Optimizing Causal Information Flow in Graph Neural Networks (AAAI 2024 · 13 citations) | |
| add pre-training | Cora · accuracy | 74.63 ±2.93 -> 88.74±0.6 | Helped | Transitivity-Preserving Graph Representation Learning for Bridging Local Connectivity and Role-Based Similarity (AAAI 2024 · 12 citations) |
| remove cross-level matching | MUTAG · accuracy (%) | 91.20 -> 90.80 | Hurt | Empowering Dual-Level Graph Self-Supervised Pretraining with Motif Discovery (AAAI 2024 · 11 citations) |
| Add self-expressive generator | Texas, OOD attack · accuracy | 62.17 -> 71.15 | Helped | Refining Latent Homophilic Structures over Heterophilic Graphs for Robust Graph Convolution Networks (AAAI 2024 · 11 citations) |
| replace vanilla global attention with BGA | Cora · test accuracy | 54.86 -> 68.72 | Helped | Less is More: on the Over-Globalizing Problem in Graph Transformers (ICML 2024 · 9 citations) |
| add the multiview kernel to GCKM | Chameleon · test accuracy (%) | 30.17 -> 38.15 | Helped | Unsupervised Neighborhood Propagation Kernel Layers for Semi-supervised Node Classification (AAAI 2024 · 6 citations) |
| Add confidence scores to the type-conditioned processing | IMDB (20%, 40%, and 80% Training) · Macro-F1 | Helped | Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach (AAAI 2025 · 5 citations) | |
| add graph-based flow representations | NF-CSE-CIC-IDS2018-v3 · PR-AUC | 0.7398 -> 0.8819 | Helped | Self-Supervised Learning of Graph Representations for Network Intrusion Detection (NeurIPS 2025 · 4 citations) |
| Use attention that combines node-level and cluster-level information instead of cluster-level-only attention | Graph classification datasets in the attention ablation · performance | Helped | Cluster-wise Graph Transformer with Dual-granularity Kernelized Attention (NeurIPS 2024 · 4 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 |
|---|---|---|---|---|
| NodeFormer | CiteSeer | accuracy | 72.5 – 76.33 | 2 papers, e.g. NeurIPS 2023 and NeurIPS 2023 |
What reviewers will expect
Datasets and baselines named in the 32 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 · 10 of 32 papers
- Citeseer · 10 of 32 papers
- PubMed · 8 of 32 papers
- Photo · 5 of 32 papers
- WikiCS · 4 of 32 papers
- Squirrel · 4 of 32 papers
- MUTAG · 3 of 32 papers
- PROTEINS · 3 of 32 papers
Baselines
- GCN · 7 of 32 papers
- GAT · 6 of 32 papers
- GraphMAE · 3 of 32 papers
- GIN · 3 of 32 papers
- GPRGNN · 3 of 32 papers
- SGFormer · 3 of 32 papers
- TAPE · 2 of 32 papers
- GraphSAGE · 2 of 32 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 111 papers; from OpenAlex and the papers' affiliations).
- Nanyang Technological University · 6 papers
- Tsinghua University · 4 papers
- Beijing University of Posts and Telecommunications · 4 papers
- Korea Advanced Institute of Science and Technology · 4 papers
- The Chinese University of Hong Kong · 4 papers
- Beihang University · 4 papers
- Shanghai Jiao Tong University · 4 papers
- Nanjing University of Science and Technology · 3 papers
- University of Chinese Academy of Sciences · 3 papers
- Northeastern University · 3 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
- Heterophily and imbalance on graphs (178 papers)
- LLMs and graph foundation models (117 papers)
- Brain and biological graph learning (78 papers)
- Multi-view graph clustering (70 papers)
Elsewhere in the topic
- Diffusion-based graph generation (138 papers)
- Graph anomaly detection (101 papers)
- Molecular graph pretraining (101 papers)
- Graph domain adaptation (82 papers)
- Graph contrastive learning analysis (61 papers)
- Graph out-of-distribution generalization (58 papers)
- Contrastive learning with negatives and views (51 papers)
- Augmentation-free graph learning (25 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).