:sparkles: GOOD: A Graph Out-of-Distribution Benchmark :sparkles:

Documentation | NeurIPS 2022 Paper | Preprint

This repo maintains and updates GOOD benchmark which is accepted by NeurIPS 2022 Datasets and Benchmarks Track. :smile:

News

• Algorithm GIL added: Learning Invariant Graph Representations for Out-of-Distribution Generalization (NeurIPS 2022) [Mar 11th, 2024]
• Our new graph OOD work on graph-level tasks: Joint Learning of Label and Environment Causal Independence for Graph Out-of-Distribution Generalization (NeurIPS 2023).

Roadmap

Tutorial

• More detailed tutorial to add new algorithms. Please refer to Add a new algorithm.

Algorithms

* denotes the method is reproduced by its authors.

• [Beta: feedback is welcome] Learning Invariant Graph Representations for Out-of-Distribution Generalization
• Learning Causally Invariant Representations for Out-of-Distribution Generalization on Graphs [the official implementation]*
• Interpretable and Generalizable Graph Learning via Stochastic Attention Mechanism

Datasets

We are planning to include more graph out-of-distribution datasets for your convenience.

• Twitter from this survey, GOOD style splits shared by LECI.
• Parts of DrugOOD (Task: LBAP, Noise level: core)

Features

• Updated final result output for an easier result gathering. [Feb 20th updates]

Leaderboard [Feb 20th updates]

• The leaderboard 1.1.0 on latest datasets will have larger hyperparameter spaces and more runs for hyperparameter sweeping.
• Results will be posted on this leaderboard gradually.

Table of contents

• Overview
• Why GOOD?
• Installation
• Quick tutorial
• Add a new algorithm
• Citing GOOD
• License
• Contact

Overview

GOOD (Graph OOD) is a graph out-of-distribution (OOD) algorithm benchmarking library depending on PyTorch and PyG to make develop and benchmark OOD algorithms easily.

Currently, GOOD contains 11 datasets with 17 domain selections. When combined with covariate, concept, and no shifts, we obtain 51 different splits. We provide performance results on 12 commonly used baseline methods (ERM, IRM, VREx, GroupDRO, Coral, DANN, MixupForGraph, DIR, GSAT, CIGA, EERM,SRGNN) including 6 graph specific methods with 10 random runs.

The GOOD dataset summaries are shown in the following figure.

Why GOOD?

Whether you are an experienced researcher of graph out-of-distribution problems or a first-time learner of graph deep learning, here are several reasons to use GOOD as your Graph OOD research, study, and development toolkit.

• Easy-to-use APIs: GOOD provides simple APIs for loading OOD algorithms, graph neural networks, and datasets so that you can take only several lines of code to start.
• Flexibility: Full OOD split generalization code is provided for extensions and any new graph OOD dataset contributions. OOD algorithm base class can be easily overwritten to create new OOD methods.
• Easy-to-extend architecture: In addition to playing as a package, GOOD is also an integrated and well-organized project ready to be further developed. All algorithms, models, and datasets can be easily registered by register and automatically embedded into the designed pipeline like a breeze! The only thing the user needs to do is write your own OOD algorithm class, your own model class, or your new dataset class. Then you can compare your results with the leaderboard.
• Easy comparisons with the leaderboard: We provide insightful comparisons from multiple perspectives. Any research and studies can use our leaderboard results for comparison. Note that this is a growing project, so we will include new OOD algorithms gradually. Besides, if you hope to include your algorithms in the leaderboard, please contact us or contribute to this project. A big welcome!

Installation

• Unbuntu >= 18.04

Conda dependencies

GOOD depends on PyTorch (>=1.6.0), PyG (>=2.0), and RDKit (>=2020.09.5). For more details: conda environment

Note that we currently test on PyTorch (==1.10.1), PyG (==2.0.4), RDKit (==2020.09.5); thus we strongly encourage to install these versions.

Warning: Please install with cuda >= 11.3 to avoid unexpected cuda errors.

Recommended installation examples:

• PyTorch 1.10.1, PyG 2.0.4, RDKit 2020.09.5, CUDA 11.3

# Create your own conda environment, then...
conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge
conda install pyg -c pyg
conda install -c conda-forge rdkit==2020.09.5

• PyTorch 2.1.2, PyG 2.5.0, RDKit 2020.09.5, CUDA 11.8

conda install -y pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=11.8 -c pytorch -c nvidia
conda install -y  pyg -c pyg
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu118.html
conda install -c conda-forge rdkit==2020.09.5 # If newer version is used, the dataset generation may fail or results will be different.

Pip

Installation for Project usages (recommended)

git clone https://github.com/divelab/GOOD.git && cd GOOD
pip install -e .

Quick Tutorial

Run an algorithm

It is a good beginning to make it work directly. Here, we provide the CLI goodtg (GOOD to go) to access the main function located at GOOD.kernel.main:goodtg. Choosing a config file in configs/GOOD_configs, we can start a task:

goodtg --config_path GOOD_configs/GOODCMNIST/color/concept/DANN.yaml

Hyperparameter sweeping

To perform automatic hyperparameter sweeping and job launching, you can use goodtl (GOOD to launch):

goodtl --sweep_root sweep_configs --launcher MultiLauncher --allow_datasets GOODMotif --allow_domains basis --allow_shifts covariate --allow_algs GSAT --allow_devices 0 1 2 3

• --sweep_root is a config fold located at configs/sweep_configs, where we provide a GSAT algorithm hyperparameter sweeping setting example (on GOODMotif dataset, basis domain, and covariate shift).
  • Each hyperparameter searching range is specified by a list of values. Example
  • These hyperparameter configs will be transformed to be CLI argument combinations.
  • Note that hyperparameters in inner config files will overwrite the outer ones.
• --launcher denotes the chosen job launcher. Available launchers:
  • Launcher: Dummy launcher, only print.
  • SingleLauncher: Sequential job launcher. Choose the first device in --allow_devices.
  • MultiLauncher: Multi-gpu job launcher. Launch on all gpus specified by --allow_devices.
• --allow_XXX denotes the job scale. Note that for each "allow" combination (e.g. GSAT GOODMotif basis covariate), there should be a corresponding sweeping config: GSAT/GOODMotif/basis/covaraite/base.yaml in the fold specified by --sweep_root.
• --allow_devices specifies the gpu devices used to launch jobs.

Sweeping result collection and config update.

To harvest all fruits you have grown (collect all results you have run), please use goodtl with a special launcher HarvestLauncher:

goodtl --sweep_root sweep_configs --final_root final_configs --launcher HarvestLauncher --allow_datasets GOODMotif --allow_domains basis --allow_shifts covariate --allow_algs GSAT

• --sweep_root: We still need it to specify the experiments that can be harvested.
• --final_root: A config store place that will store the best config settings. We will update the best configurations (according to the sweeping) into the config files in it.

(Experimental function.)

The output numpy array:

• Rows: In-distribution train/In-distribution test/Out-of-distribution train/Out-of-distribution test/Out-of-distribution validation
• Columns: Mean/Std.

Final runs

It is sometimes not practical to run 10 rounds for hyperparameter sweeping, especially when the searching space is huge. Therefore, we can generally run hyperparameter sweeping for 2~3 rounds, then perform all rounds after selecting the best hyperparameters. Now, remove the --sweep_root, set --config_root to your updated best config saving location, and set the --allow_rounds.

goodtl --config_root final_configs --launcher MultiLauncher --allow_datasets GOODMotif --allow_domains basis --allow_shifts covariate --allow_algs GSAT --allow_devices 0 1 2 3 --allow_rounds 1 2 3 4 5 6 7 8 9 10

Note that the results are valid only after 3+ rounds experiments in this benchmark.

Final result collection

goodtl --config_root final_configs --launcher HarvestLauncher --allow_datasets GOODMotif --allow_domains basis --allow_shifts covariate --allow_algs GSAT --allow_rounds 1 2 3 4 5 6 7 8 9 10

Output: Markdown format table. (This table is also saved in the file: <Project_root>/result_table.md).

You can customize your own launcher at GOOD/kernel/launchers/.

Add a new algorithm

Please follow this documentation to add a new algorithm.

Any contributions are welcomed! Please refer to contributing for adding your algorithm into GOOD.

Leaderboard

The initial leaderboard results are listed in the paper. And the validation of these results is described here.

Leaderboard 1.1.0 with updated datasets will be available here.

Citing GOOD

If you find this repository helpful, please cite our paper.

@inproceedings{
gui2022good,
title={{GOOD}: A Graph Out-of-Distribution Benchmark},
author={Shurui Gui and Xiner Li and Limei Wang and Shuiwang Ji},
booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
year={2022},
url={https://openreview.net/forum?id=8hHg-zs_p-h}
}

License

The GOOD datasets are under MIT license. The GOOD code are under GPLv3 license.

Discussion

Please submit new issues or start a new discussion for any technical or other questions.

Contact

Please feel free to contact Shurui Gui, Xiner Li, or Shuiwang Ji!