GrokFormer

A PyTorch implementation of GrokFormer "GrokFormer: Graph Fourier Kolmogorov-Arnold Transformer".

Environment Settings

This implementation is based on Python3. To run the code, you need the following dependencies:

• torch==1.8.1
• torch-geometric==1.7.2
• scipy==1.2.1
• numpy==1.19.5
• tqdm==4.59.0
• seaborn==0.11.2
• scikit-learn==0.24.2

Node Classification Datasets

The dataset contains six homophilic benchmark datasets(Cora, Citeseer, Pubmed, Photo, Physics, WikiCS), and five heterophilic datasets(Penn94, Chameleon, Squirrel, Actor, Texas). We use the same experimental setting (60\%/20\%/20\% random splits for train/validation/test (fixed splits for Penn94) with the same epochs, run ten times (five times for Penn94), early stopping) as Specformer.

Run node classification experiment:

$ sh run.sh

Examples

Training a model on the default dataset.

Baselines links

• H2GCN
• HopGNN
• GPRGNN
• BernNet
• JacobiConv
• HiGCN
• NodeFormer
• SGFormer
• NAGphormer
• PolyFormer
• Specformer
• The implementations of others are taken from the Pytorch Geometric library

Acknowledgements

The node-level code and filter learning code are implemented based on Specformer: Spectral Graph Neural Networks Meet Transformers. The graph-level (TUD Benchmarks) code are implemented based on the public code (https://github.com/kavehhassani/mvgrl/tree/master/graph)

📖 Citation

If you find this work useful, please cite our paper:

```bibtex @inproceedings{ai2024grokformer, title = {GrokFormer: Graph Fourier Kolmogorov-Arnold Transformers}, author = {Ai, Guoguo and Pang, Guansong and Qiao, Hezhe and Gao, Yuan and Yan, Hui}, booktitle = {Proceedings of the 42st International Conference on Machine Learning (ICML)}, year = {2025} }