README.md

GOAT: A Global Transformer on Large-scale Graphs

This is the official repository for our paper GOAT: A Global Transformer on Large-scale Graphs, accepted by ICML 2023.

TL;DR: GOAT is a scalable global transformer working on large-scale homophilious & heterophilious graphs with millions of nodes.

Requirements

To fulfill the environment requirements to reproduce the results of GOAT, run the following script.

conda create --name goat python=3.8 --file requirements.txt
conda activate goat

Data downloading

Follow OGB repo for dataset downloading of ogbn-arxiv and ogbn-products. Mean-while follow LINKX repo for arxiv-year and snap-patents.

Or you can run the command below for automatic data downloading for the specific dataset:

python arxiv_ERM_ns.py \
    --dataset [dataset name] \
    --conv_type local \
    --data_root [ogb data downloading root path] \
    --linkx_data_root [linkx data downloading root path] \
    --data_downloading_flag

Experiments

Below are the command lines to reproduce the experimental results of GOAT.

Positional encodings

The full GOAT model requires precomputed positional encodings for all of the nodes within the graph. In this work we adopt the node2vec algorithm to compute such encodings. Refer to pos_enc folder for the script to computation and saving of node2vec encodings for various datasets. Don't forget to modify data loading and encoding saving path to maintain correctness.

Training GOAT

For the experimental results on dataset ogbn-arxiv:

python arxiv_ERM_ns.py \
    --dataset ogbn-arxiv \
    --lr 1e-3 \
    --batch_size 1024 \
    --test_batch_size 256 \
    --hidden_dim 128 \
    --test_freq 1 \
    --num_workers 4 \
    --conv_type full \
    --num_heads 4 \
    --num_centroids 4096 \
    --data_root [ogb data downloading root path] \
    --linkx_data_root [linkx data downloading root path]

For the experimental results on dataset ogbn-products:

python arxiv_ERM_ns.py \
    --dataset ogbn-products \
    --lr 1e-3 \
    --batch_size 512 \
    --test_batch_size 256 \
    --hidden_dim 256 \
    --test_freq 5 \
    --num_workers 4 \
    --conv_type full \
    --num_heads 2 \
    --num_centroids 4096 \
    --data_root [ogb data downloading root path] \
    --linkx_data_root [linkx data downloading root path]

For the experimental results on dataset arxiv-year, you need to mannually specify the hetero_train_prop argument. To reproduce results, select one out of [0.1, 0.2, 0.5]:

python arxiv_ERM_ns.py \
    --dataset arxiv-year \
    --lr 1e-3 \
    --batch_size 1024 \
    --test_batch_size 256 \
    --hidden_dim 128 \
    --test_freq 5 \
    --num_workers 4 \
    --conv_type full \
    --num_heads 4 \
    --num_centroids 4096 \
    --hetero_train_prop [heterophilious train proportion] \
    --data_root [ogb data downloading root path] \
    --linkx_data_root [linkx data downloading root path]

For the experimental results on dataset snap-patents, you need to mannually specify the hetero_train_prop argument. To reproduce results, select one out of [0.1, 0.2, 0.5]:

python arxiv_ERM_ns.py \
    --dataset snap-patents \
    --lr 1e-3 \
    --batch_size 8192 \
    --test_batch_size 8192 \
    --hidden_dim 128 \
    --test_freq 5 \
    --num_workers 4 \
    --conv_type full \
    --num_heads 2 \
    --num_centroids 4096 \
    --hetero_train_prop [heterophilious train proportion] \
    --data_root [ogb data downloading root path] \
    --linkx_data_root [linkx data downloading root path]

Citing GOAT

If you find GOAT useful, please cite our paper.

@article{kong2023goat,
  title={GOAT: A Global Transformer on Large-scale Graphs},
  author={Kong, Kezhi and Chen, Jiuhai and Kirchenbauer, John and Ni, Renkun and Bruss, C Bayan and Goldstein, Tom},
  year={2023}
}