GNN architectures: research map
1,634 papers on GNN architectures in Graph learning, from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR (2016–2026), grouped into 4 clusters and 17 approaches. The busiest year so far is 2025.
What it is
Graph neural networks learn on graph data by passing messages between neighbouring nodes. This topic asks which architectures are most expressive (can tell more graphs apart), why deep GNNs blur node features (oversmoothing) and how to make GNNs robust, explainable and equivariant.
Terms you will see on this page:
- Message passing
- each node updating its features from its neighbours' features.
- Weisfeiler-Leman (WL) test
- a classic graph test used to measure what GNNs can distinguish.
- Oversmoothing
- node features becoming indistinguishable as layers are stacked.
Start reading with the most cited papers here:
- Graph Attention Networks (ICLR 2018 · 9,105 citations)
- Neural Message Passing for Quantum Chemistry (ICML 2017 · 3,844 citations)
Within Graph learning, its share shrank from 37.3% in 2023–24 to 32.2% in 2025–26 (474 → 508 papers at ICML, NeurIPS, ICLR, CVPR and AAAI, the venues with data for all four years).
Explore GNN architectures in the interactive map
Working on something in this topic? Describe your idea in scime atlas to see which approach it falls under, the closest papers by meaning and how crowded the spot has become.
What changed recently
Approaches whose share of the topic moved most between 2023–24 and 2025–26 (ICML, NeurIPS, ICLR, CVPR and AAAI):
- GNN generalization theory rising: 11% → 15% of the topic (53 → 75 papers)
- Weisfeiler-Leman expressivity hierarchies rising: 4% → 6% of the topic (20 → 32 papers)
- Graph structure learning rising: 13% → 15% of the topic (61 → 75 papers)
- Oversmoothing and heterophily falling: 5% → 2% of the topic (24 → 12 papers)
- Simple and scalable GNNs falling: 9% → 7% of the topic (45 → 35 papers)
Papers from 2025–2026 that are already cited most:
- When Hypergraph Meets Heterophily: New Benchmark Datasets and Baseline (AAAI 2025 · 34 citations)
- Mind Individual Information! Principal Graph Learning for Multimedia Recommendation (AAAI 2025 · 25 citations)
- Modality-Independent Graph Neural Networks with Global Transformers for Multimodal Recommendation (AAAI 2025 · 20 citations)
- DeepSN: A Sheaf Neural Framework for Influence Maximization (AAAI 2025 · 10 citations)
Approaches and key papers
GNN theory and trustworthiness · 851 papers
Keywords: gnn, explanations, trained
Approaches in this cluster:
- GNN generalization theory (200 papers)
Analyze size generalization and statistical properties of GNNs, with benchmarks. - Simple and scalable GNNs (154 papers)
Replace or simplify GNNs with label propagation, linear models, and subsampling. - Robustness, privacy, and fairness of GNNs (97 papers)
Defend GNNs against attacks and study their privacy and fairness. - Physics-inspired GNNs (86 papers)
Apply equivariant and diffusion-based GNNs to physical simulation. - GNN explainability (73 papers)
Evaluate and generate explanations for GNN predictions. - Graph structure learning (241 papers)
Learn or refine graph structure to improve node classification.
Most cited and most cited since 2024:
- GNNExplainer: Generating Explanations for Graph Neural Networks (NeurIPS 2019 · 687 citations)
- Do Transformers Really Perform Badly for Graph Representation? (NeurIPS 2021 · 378 citations)
- GreedyViG: Dynamic Axial Graph Construction for Efficient Vision GNNs (CVPR 2024 · 39 citations)
- Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024 · 33 citations)
Message passing and equivariance · 307 papers
Keywords: message passing, equivariant, mpnns
Approaches in this cluster:
- Expressive equivariant message passing (115 papers)
Analyze limits of message passing and build more powerful equivariant or subgraph GNNs. - Graph rewiring and scalable message passing (75 papers)
Rewire graphs and add virtual nodes to improve message passing. - Structural GNN variants (71 papers)
Develop new graph architectures using geometry, trees, templates, and commute-time structure. - Equivariant networks for molecules (46 papers)
Predict molecular and crystal properties with equivariant message passing.
Most cited and most cited since 2024:
- Neural Message Passing for Quantum Chemistry (ICML 2017 · 3,844 citations)
- Learning to Simulate Complex Physics with Graph Networks (ICML 2020 · 434 citations)
- TextGT: A Double-View Graph Transformer on Text for Aspect-Based Sentiment Analysis (AAAI 2024 · 49 citations)
- Provably Powerful Graph Neural Networks for Directed Multigraphs (AAAI 2024 · 28 citations)
Graph convolutional networks · 303 papers
Keywords: convolutional, gcns, spectral
Approaches in this cluster:
- Deep graph convolution and simplification (101 papers)
Make GCNs deeper and simpler with spectral analysis. - Oversmoothing and heterophily (63 papers)
Analyze and mitigate oversmoothing and heterophily in GNNs. - Spectral GNN design (47 papers)
Improve spectral GNN filters and understand their reliance on Fourier basis. - Theory of graph convolutional networks (92 papers)
Analyze convergence and transferability of graph convolutional networks on random graphs.
Most cited and most cited since 2024:
- Graph Attention Networks (ICLR 2018 · 9,105 citations)
- Simplifying Graph Convolutional Networks (ICML 2019 · 1,272 citations)
- PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · 45 citations)
- GCNext: Towards the Unity of Graph Convolutions for Human Motion Prediction (AAAI 2024 · 28 citations)
GNN expressivity · 173 papers
Keywords: wl, expressive, power
Approaches in this cluster:
- Weisfeiler-Leman expressivity hierarchies (66 papers)
Characterize GNN expressive power with WL-style hierarchies. - Subgraph and distance-aware GNNs (51 papers)
Boost GNN expressivity with identity, distance encoding, and subgraph information. - Substructure-aware GNN expressivity (56 papers)
Analyze GNN capability to count motifs and substructures, including topological and geometric variants.
Most cited and most cited since 2024:
- How Powerful are Graph Neural Networks? (ICLR 2019 · 374 citations)
- Identity-aware Graph Neural Networks (AAAI 2021 · 140 citations)
- Union Subgraph Neural Networks (AAAI 2024 · 11 citations)
- Weisfeiler and Lehman Go Paths: Learning Topological Features via Path Complexes (AAAI 2024 · 7 citations)
