migalkin / RWL

Weisfeiler and Leman Go Relational (LOG 2022)

Weisfeiler and Leman Go Relational (LOG 2022)

This is the official code base of the paper

Weisfeiler and Leman Go Relational

Pablo Barcelo, Mikhail Galkin, Christopher Morris, Miguel Romero Orth

Overview

This repo contains the code for reproducing the experiments on R-GCN and CompGCN with the one-hot feature initialization strategy.

Notice on the k-RN architecture: We plan to update the repo with the k-RN implementation as soon as we come up with the meaningful relational dataset to evaluate k-RNs.

Installation

The experiments were performed on Python 3.8.

Dependencies:

torch                 1.10.0
torch-cluster         1.5.9
torch-geometric       2.0.2
torch-scatter         2.0.9
torch-sparse          0.6.12

Optionally, install wandb for results tracking, prepend WANDB_ENTITY=yourentity to the running script and use the --wandb flag.

• To run experiments on the AIFB dataset with the fast version of R-GCN and 4d features:

python main.py --dataset AIFB --lr 0.001 --epochs 8001 --rgcn_fast --drop_bias --dim 4

• For the modified R-GCN with the additional MLP over aggregated node features:

python main.py --dataset AIFB --lr 0.001 --epochs 8001 --rgcn_fast --mod_rgcn --drop_bias --dim 4

• For CompGCN:

python main.py --dataset AIFB --lr 0.001 --epochs 8001 --compgcn --dim 4

• For CompGCN without directional updates:

python main.py --dataset AIFB --lr 0.001 --epochs 8001 --compgcn --dim 4 --compgcn_no_dir

• For CompGCN without relation updates:

python main.py --dataset AIFB --lr 0.001 --epochs 8001 --compgcn --dim 4 --compgcn_no_relupd

• For CompGCN without adjacency normalization:

python main.py --dataset AIFB --lr 0.001 --epochs 8001 --compgcn --dim 4 --no_norm

You can combine those flags for CompGCN as well.

Experiments on the big AM dataset are forced on a CPU due to the dataset size.

Options for --msg_func: transe, distmult, rotate

Options for --aggr_func: add, mean

Please refer to the Appendix F in the paper for the full set of hyperparameters.

Citation

If you find this project useful in your research, please cite the following paper

@inproceedings{
    barcelo2022weisfeiler,
    title={Weisfeiler and Leman Go Relational},
    author={Pablo Barcelo and Mikhail Galkin and Christopher Morris and Miguel Romero Orth},
    booktitle={Learning on Graphs Conference},
    year={2022},
    url={https://openreview.net/forum?id=wY_IYhh6pqj}
}