The official code of the paper Beyond Weisfeiler-Lehman: A Quantitative Framework for GNN Expressiveness.
Required packages:
- torch 1.9.0
- torch-scatter 2.0.8
- torch-geometric 1.7.2
The dataset for the counting task is generated by running the command python data/count.py. The training commands are contained in the bash scripts main.zinc.sh (for both ZINC-subset and ZINC-full datasets), main.alchemy.sh (for the Alchemy dataset), and main.count.sh (for both homomorphism and cycle counting tasks).
Log files in folders named with the prefix result contains training results for different model configurations. In each file, the i-th row contains metrics at the i-th training epoch, which is in the following format:
| Learning rate | Training loss | Validation set MAE | Test set MAE |
|---|
The file names begin with a prefix indicating the model type: MP for message-passing NN, Sub for subgraph GNN, L for local 2-GNN, and LF for local 2-FGNN. The -G suffix denotes the inclusion of global aggregation in each layer.
The file name follows the format: f"result.zinc/{model}-{subset}-{max_dis}-{num_layer}x{dim_embed}-{bs}-{lr}-{seed}.txt", where each component represents the model string, subset or full dataset, maximum distance encoding, batch size, learning rate, and random seed.
The file name follows the format: f"result.alchemy/{model}-{max_dis}-{num_layer}x{dim_embed}-{bs}-{lr}-{seed}.txt", where each component represents the model string, maximum distance encoding, batch size, learning rate, and random seed.
The file name follows the format: f"result.count/{model}-{task}-{max_dis}-{num_layer}x{dim_embed}-{bs}-{lr}-{seed}.txt", where each component represents the model string, task name, maximum distance encoding, batch size, learning rate, and random seed. The task name is a tuple of structure name (see dataset.py for detailed definitions) and counting level (g for graph level, v for node level, and e for edge level).
- 2024.01.14: release code.
- 2024.01.19: update citation.
Our code is based on SWL.
If you find this repository useful, please cite our work:
@article{zhang2024beyond,
title={Beyond Weisfeiler-Lehman: A Quantitative Framework for GNN Expressiveness},
author={Zhang, Bohang and Gai, Jingchu and Du, Yiheng and Ye, Qiwei and He, Di and Wang, Liwei},
journal={arXiv preprint arXiv:2401.08514},
year={2024}
}