Research map / Graph learning / GNN architectures
Oversmoothing and heterophily
Analyze and mitigate oversmoothing and heterophily in GNNs.
60 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in GNN architectures, keyword cluster “convolutional · gcns · spectral”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Leveling off Several papers from 2025–26 still report on 3 tracked benchmarks, but the best result has barely moved since 2024.
24 papers in 2023–24 and 12 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 5.1% to 2.4%.
Open questions the data points to
- Papers report GCN on Squirrel (accuracy) anywhere from 23.96 to 36.9: comparisons against it depend on whose number you take.
- Papers report MLP on Actor (accuracy (%)) anywhere from 35 to 39.68: comparisons against it depend on whose number you take.
What looks settled
- Cora (accuracy): best up to 2024 90.02, best in 2025–26 88.07 from 2 papers, at most 0.2% better, if at all.
- Citeseer (accuracy): best up to 2024 81.76, best in 2025–26 78.09 from 2 papers, at most 0.2% better, if at all.
- PubMed (accuracy): best up to 2024 91.3, best in 2025–26 89.64 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
- 2018 · Deep Learning and Quantum Entanglement: Fundamental Connections with Implications to Network Design (ICLR 2018 · 35 citations)
- 2019 · Predict then Propagate: Graph Neural Networks meet Personalized PageRank (ICLR 2019 · 428 citations)
- 2020 · PairNorm: Tackling Oversmoothing in GNNs (ICLR 2020 · 225 citations)
- 2021 · GraphMix: Improved Training of GNNs for Semi-Supervised Learning (AAAI 2021 · 121 citations)
- 2022 · Is Homophily a Necessity for Graph Neural Networks? (ICLR 2022 · 55 citations)
- 2023 · Demystifying Oversmoothing in Attention-Based Graph Neural Networks (NeurIPS 2023 · 10 citations)
- 2024 · PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · 45 citations)
- 2025 · Deep Hypergraph Neural Networks with Tight Framelets (AAAI 2025 · 15 citations)
- 2026 · MORGAN: To Bridge Mixture of Experts and Spectral Graph Neural Network (AAAI 2026 · 1 citation)
Earliest papers here
- Deep Learning and Quantum Entanglement: Fundamental Connections with Implications to Network Design (ICLR 2018 · 35 citations)
- Generalized Value Iteration Networks:Life Beyond Lattices (AAAI 2018 · 21 citations)
- Predict then Propagate: Graph Neural Networks meet Personalized PageRank (ICLR 2019 · 428 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)
8 papers, 2022–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 87.12 | Is Homophily a Necessity for Graph Neural Networks? (ICLR 2022 · 55 citations) |
| 2023 | 87.99 | Graph Positional Encoding via Random Feature Propagation (ICML 2023 · 0 citations) |
| 2024 | 90.02 | PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · 45 citations) |
Citeseer · accuracy (higher is better)
7 papers, 2022–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 76.5 | Is Homophily a Necessity for Graph Neural Networks? (ICLR 2022 · 55 citations) |
| 2024 | 81.76 | PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · 45 citations) |
Texas · accuracy (higher is better)
7 papers, 2022–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 87.96 | $p$-Laplacian Based Graph Neural Networks (ICML 2022 · 4 citations) |
| 2024 | 95.57 | PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · 45 citations) |
Actor · accuracy (higher is better)
6 papers, 2022–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 40.62 | $p$-Laplacian Based Graph Neural Networks (ICML 2022 · 4 citations) |
| 2024 | 42.72 | PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · 45 citations) |
PubMed · accuracy (higher is better)
6 papers, 2022–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 88.52 | Is Homophily a Necessity for Graph Neural Networks? (ICLR 2022 · 55 citations) |
| 2024 | 91.3 | PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · 45 citations) |
Wisconsin · accuracy (higher is better)
6 papers, 2022–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 95.37 | $p$-Laplacian Based Graph Neural Networks (ICML 2022 · 4 citations) |
What has been tried
52 papers here report 504 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 9%, hurt 16%, 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 |
|---|---|---|---|---|
| use generalized Laplacian normalization instead of traditional Laplacian normalization (η=0.5) | Cora, CiteSeer, PubMed, Texas, Actor, Cornell · accuracy | 82.33_{±\text{0.83}}, 69.01_{±\text{1.08}}, 79.73 _{±\text{0.97}}, 57.55_{±\text{2.66}}, 33.28_{±\tex | Helped | PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · 45 citations) |
| add high-pass component | Cora · accuracy | 80.34 -> 81.51 | Helped | Deep Hypergraph Neural Networks with Tight Framelets (AAAI 2025 · 15 citations) |
| Add the low-frequency GCN filter and two-stage attention | Appliances, Toys, Grocery, Home · NDCG | 0.2457 -> 0.4377; 0.2236 -> 0.5501; 0.2560 -> 0.5959; 0.2375 -> 0.6610 | Helped | Spectral-Based Graph Neural Networks for Complementary Item Recommendation (AAAI 2024 · 15 citations) |
| Remove the E1 evolutionary prompt | PubMed · accuracy | 85.13±0.26 -> 84.88±0.43 | Hurt | AutoSGNN: Automatic Propagation Mechanism Discovery for Spectral Graph Neural Networks (AAAI 2025 · 9 citations) |
| vary hop size from k=2 to k=6 | Cora · accuracy | 82.46 -> 84.38 | Helped | SEA-GWNN: Simple and Effective Adaptive Graph Wavelet Neural Network (AAAI 2024 · 6 citations) |
| Add edge-transmitted signals alongside node signals instead of using node signals alone | Cora, p_r=20% · accuracy | 63.54 -> 74.81 | Helped | Towards Inductive Robustness: Distilling and Fostering Wave-Induced Resonance in Transductive GCNs against Graph Adversarial Attacks (AAAI 2024 · 5 citations) |
| Add fairness regularization (λ_fair=4.0) | NBA · average within-group unfairness | 0.013 ± 0.005 -> 0.000 ± 0.001 | Helped | Networked Inequality: Preferential Attachment Bias in Graph Neural Network Link Prediction (ICML 2024 · 5 citations) |
| Vary the confidence threshold τ from 0.7 to 1 | Citeseer · Accuracy | Mixed | Integrating Co-Training with Edge Discrimination to Enhance Graph Neural Networks Under Heterophily (AAAI 2025 · 1 citation) | |
| Increase GNN depth from two to four layers | Coauthor-CS and other evaluated datasets · Accuracy-curve spectral tendency | No clear effect | Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · 1 citation) | |
| increase the averaged node degree | Synthetic HSBM · accuracy | Helped | Understanding Heterophily for Graph Neural Networks (ICML 2024 · 1 citation) | |
| Increase attention intensity when structure noise is high and feature noise is low | CSBM node classification · SNR | \mu'/\sigma' = \sqrt{np} \cdot \mu/\sigma | Helped | Graph Attention is Not Always Beneficial: A Theoretical Analysis of Graph Attention Mechanisms via Contextual Stochastic Block Models (ICML 2025 · 0 citations) |
| Use K=1 rather than K=10 power iterations | PROTEINS · attack success rate | 19.89 -> 21.04 | Hurt | Enhancing Graph Classification Robustness with Singular Pooling (NeurIPS 2025 · 0 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 | Squirrel | accuracy | 23.96 – 36.9 | 2 papers, e.g. ICML 2023 and AAAI 2023 |
| MLP | Actor | accuracy (%) | 35 – 39.68 | 2 papers, e.g. ICML 2024 and ICML 2022 |
| MLP | Chameleon | accuracy (%) | 48.02 – 54.29 | 2 papers, e.g. ICML 2022 and ICML 2024 |
| GAT | Citeseer | accuracy (%) | 72.6 – 80.52 | 2 papers, e.g. ICLR 2023 and AAAI 2024 |
| MLP | Squirrel | accuracy (%) | 33.8 – 36.63 | 2 papers, e.g. ICML 2022 and ICML 2024 |
| MLP | Cornell | accuracy | 80.3 – 84.86 | 2 papers, e.g. ICML 2022 and ICLR 2022 |
What reviewers will expect
Datasets and baselines named in the 9 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 · 5 of 9 papers
- CiteSeer · 4 of 9 papers
- PubMed · 4 of 9 papers
- Chameleon · 2 of 9 papers
- Squirrel · 2 of 9 papers
- Texas · 2 of 9 papers
- Cornell · 2 of 9 papers
- Wisconsin · 2 of 9 papers
Baselines
- GCN · 4 of 9 papers
- GAT · 3 of 9 papers
- APPNP · 2 of 9 papers
- ChebNet · 2 of 9 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 36 papers; from OpenAlex and the papers' affiliations).
- University of Science and Technology of China · 3 papers
- University of Electronic Science and Technology of China · 2 papers
- Massachusetts Institute of Technology · 2 papers
- KTH Royal Institute of Technology · 2 papers
- Ecole Polytechique · 2 papers
- NVIDIA · 2 papers
- Ben-Gurion University of the Negev · 2 papers
- Imperial College London · 2 papers
- Tsinghua University · 2 papers
Use it
Download BibTeX of 36 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
- Deep graph convolution and simplification (82 papers)
- Spectral GNN design (37 papers)
- Theory of graph convolutional networks (83 papers)
Elsewhere in the topic
- Graph structure learning (201 papers)
- GNN generalization theory (187 papers)
- Simple and scalable GNNs (131 papers)
- Expressive equivariant message passing (109 papers)
- Robustness, privacy, and fairness of GNNs (94 papers)
- Physics-inspired GNNs (81 papers)
- Graph rewiring and scalable message passing (72 papers)
- Weisfeiler-Leman expressivity hierarchies (66 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).