GNNBenchmark
MNIST
Structure
70,000Graphs
71Nodes
565Edges
10Classes
3Node Feature Dim
1Edge Feature Dim
Coverage
173Results
1Variants
34Papers
- Beyond Oversquashing: Understanding Signal Propagation in GNNs Via Observables Beyond Oversquashing: Understanding Signal Propagation in GNNs Via Observables
- GCCM: Enhancing Generative Graph Prediction via Contrastive Consistency Model GCCM: Enhancing Generative Graph Prediction via Contrastive Consistency Model
- GNN Explanations that do not Explain and How to find Them GNN Explanations that do not Explain and How to find Them
- A Hierarchical Quantized Tokenization Framework for Task-Adaptive Graph Representation Learning A Hierarchical Quantized Tokenization Framework for Task-Adaptive Graph Representation Learning
- TANGO: Graph Neural Dynamics via Learned Energy and Tangential Flows TANGO: Graph Neural Dynamics via Learned Energy and Tangential Flows
- Plain Transformers Can be Powerful Graph Learners Plain Transformers Can be Powerful Graph Learners
- TMLR 2025: Towards Efficient Training of Graph Neural Networks: A Multiscale Approach Towards Efficient Training of Graph Neural Networks: A Multiscale Approach TMLR 2025
- KDD 2024: Benchmarking Positional Encodings for GNNs and Graph Transformers Benchmarking Positional Encodings for GNNs and Graph Transformers KDD 2024
- Range-aware Positional Encoding via High-order Pretraining: Theory and Practice Range-aware Positional Encoding via High-order Pretraining: Theory and Practice
- Graph Triple Attention Network: A Decoupled Perspective Graph Triple Attention Network: A Decoupled Perspective
- Differential Encoding for Improved Representation Learning Over Graphs Differential Encoding for Improved Representation Learning Over Graphs
- Graph in Graph Neural Network Graph in Graph Neural Network
- HyperAggregation: Aggregating over Graph Edges with Hypernetworks HyperAggregation: Aggregating over Graph Edges with Hypernetworks
- Automatic Graph Topology-Aware Transformer Automatic Graph Topology-Aware Transformer
- RandAlign: A Parameter-Free Method for Regularizing Graph Convolutional Networks RandAlign: A Parameter-Free Method for Regularizing Graph Convolutional Networks
- TMLR 2024: Contextualized Messages Boost Graph Representations Contextualized Messages Boost Graph Representations TMLR 2024
- KDD 2024: Graph Mamba: Towards Learning on Graphs with State Space Models Graph Mamba: Towards Learning on Graphs with State Space Models KDD 2024
- Topology-Informed Graph Transformer Topology-Informed Graph Transformer
- ICML 2023: Recurrent Distance Filtering for Graph Representation Learning Recurrent Distance Filtering for Graph Representation Learning ICML 2023
- Generative Explanations for Graph Neural Network: Methods and Evaluations Generative Explanations for Graph Neural Network: Methods and Evaluations
- Enhancing Graph Neural Networks with Quantum Computed Encodings Enhancing Graph Neural Networks with Quantum Computed Encodings
- JMLR 2023: GNNBenchmark GNNBenchmark JMLR 2023
- ICML 2022: A Generalization of ViT/MLP-Mixer to Graphs A Generalization of ViT/MLP-Mixer to Graphs ICML 2022
- ICML 2022: Equivariance with Learned Canonicalization Functions Equivariance with Learned Canonicalization Functions ICML 2022
- ICML 2022: Going Deeper into Permutation-Sensitive Graph Neural Networks Going Deeper into Permutation-Sensitive Graph Neural Networks ICML 2022
- NeurIPS 2022: Recipe for a General, Powerful, Scalable Graph Transformer Recipe for a General, Powerful, Scalable Graph Transformer NeurIPS 2022
- ICML 2022: Graph-Coupled Oscillator Networks Graph-Coupled Oscillator Networks ICML 2022
- TMLR 2022: Rewiring with Positional Encodings for Graph Neural Networks Rewiring with Positional Encodings for Graph Neural Networks TMLR 2022
- SoGCN: Second-Order Graph Convolutional Networks SoGCN: Second-Order Graph Convolutional Networks
- Parameterized Hypercomplex Graph Neural Networks for Graph Classification Parameterized Hypercomplex Graph Neural Networks for Graph Classification
- ICLR 2021: Topological Graph Neural Networks Topological Graph Neural Networks ICLR 2021
- NeurIPS 2020: Rethinking pooling in graph neural networks Rethinking pooling in graph neural networks NeurIPS 2020
- Formatting Instructions for ICLR 2021 Conference Submissions Formatting Instructions for ICLR 2021 Conference Submissions
- Hierarchical Representation Learning in Graph Neural Networks with Node Decimation Pooling Hierarchical Representation Learning in Graph Neural Networks with Node Decimation Pooling
10-class graph classification Graph classification on MNIST superpixel graphs (10 digit classes, 70k graphs). Uses the benchmark-provided synthetic or superpixel node/edge features. Evaluated by Accuracy.
Dataset details
Task
Graph classification on MNIST superpixel graphs (10 digit classes, 70k graphs). The primary catalog metric is Accuracy.
Data
Graph classification on MNIST superpixel graphs (10 digit classes, 70k graphs).
Features
The benchmark release supplies fixed graph tensors and task-specific node or edge inputs. Image graphs use superpixel appearance and position, synthetic node tasks use generated signals, TSP uses coordinates, and CSL is intentionally featureless.
Splits and evaluation
The catalog records these protocols or variants: Standard split. Evaluation uses Accuracy (higher is better). Exact masks or folds must come from the cited release.
Quirks and pitfalls
Use the released train/validation/test files rather than resplitting. These tasks were designed to expose specific architectural limitations, and preprocessing such as changing superpixels, adding positional encodings, or balancing edge labels changes the benchmark.
Sources
SOTA Progression
new SOTA
matches SOTA at that time
other proposed result
other baseline result
Leaderboard
| # | Model | Accuracy ↑ | Paper |
|---|---|---|---|
| 1 | Schrödinger | 0.9913
±
0.0004 | |
| 2 | CNN | 0.9907
±
0.0007 | |
| 3 | MPNN | 0.9895
±
0.0006 | |
| 4 | GraphCON-GAT | 0.9891 | |
| 5 | GIG-GatedGCN | 0.9880
±
0.0003 | |
| 6 | PNCNN | 0.9876 | |
| 7 | Ours(100K) | 0.9874
±
0.0006 | |
| 8 | GraphCON-GCN | 0.9868 | |
| 9 | PPGT | 0.9861
±
0.0010 | |
| 10 | differential encoding | 0.9856
±
0.0006 | |
| 11 | GatedGCN | 0.9855
±
0.0002 | |
| 12 | GatedGCN + RandAlign | 0.9855
±
0.0002 | |
| 13 | Exphormer | 0.9855
±
0.0004 | |
| 14 | GraphCON-GIN | 0.9853 | |
| 15 | RandAlign | 0.9851
±
0.0005 | |
| 16 | LRGA + GatedGCN | 0.9847
±
0.0016 | |
| 17 | GatedGCN-ViT | 0.9846
±
0.0009 | |
| 18 | EdgeGCN | 0.9843
±
0.0006 | |
| 19 | CKGCN | 0.9842
±
0.0015 | |
| 20 | Exphormer + RRWP | 0.9842
±
0.0018 | |
| 21 | Exphormer + RWDIFF | 0.9842
±
0.0006 | |
| 22 | Exphormer + noPE | 0.9841
±
0.0005 | |
| 23 | LRGA + GAT | 0.9841
±
0.0008 | |
| 24 | Exphormer + GCKN | 0.9840
±
0.0007 | |
| 25 | Exphormer + WLPE | 0.9840
±
0.0016 | |
| 26 | GMN | 0.9839
±
0.0018 | |
| 27 | GRED | 0.9838
±
0.0001 | |
| 28 | EGTAS | 0.9837
±
0.0026 | |
| 29 | Exphormer + ESLapPE | 0.9835
±
0.0011 | |
| 30 | LRGA+GCN | 0.9834
±
0.0006 | |
| 31 | GIN | 0.9833
±
0.0011 | |
| 32 | Graph MLP-Mixer | 0.9832
±
0.0004 | |
| 33 | GatedGCN-MLP-Mixer | 0.9832
±
0.0004 | |
| 34 | GraphCON-GatedGCN | 0.9827 | |
| 35 | Exphormer + LapPE | 0.9827
±
0.0007 | |
| 36 | Exphormer + RWSE | 0.9825
±
0.0008 | |
| 37 | GatedGCN | 0.9824
±
0.0004 | |
| 38 | GCCM | 0.9824
±
0.0006 | |
| 39 | DeGTA | 0.9823
±
0.0011 | |
| 40 | TIGT | 0.9823
±
0.0013 | |
| 41 | TAU-GRIT | 0.9821
±
0.0010 | |
| 42 | GPS + Mamba | 0.9821
±
0.0004 | |
| 43 | GRIT 2-QiQRW | 0.9820
±
0.0005 | |
| 44 | GINE-ViT | 0.9820
±
0.0005 | |
| 45 | TANGO-GPS | 0.9820
±
0.0011 | |
| 46 | GraphGPS + ESLapPE | 0.9818
±
0.0012 | |
| 47 | GraphGPS + GCKN | 0.9818
±
0.0012 | |
| 48 | EGT | 0.9817
±
0.0009 | |
| 49 | GRIT (our run) | 0.9816
±
0.0005 | |
| 50 | GPS + HOPE-WavePE | 0.9815
±
0.0011 |
SOTA performance
set a new SOTA at the time
matched the SOTA at the time