Skip to content

Latest commit

 

History

History
62 lines (40 loc) · 2.26 KB

README.md

File metadata and controls

62 lines (40 loc) · 2.26 KB

MARIO: Model Agnostic Recipe for Improving OOD Generalization of Graph Contrastive Learning

Official implementation of paper
MARIO: Model Agnostic Recipe for Improving OOD Generalization of Graph Contrastive Learning

Yun Zhu, Haizhou Shi, Zhenshuo Zhang, Siliang Tang†

In WWW 2024

Our codes are built up on GOOD.

Overview

Setup

conda create -n MARIO python==3.9
conda activate MARIO 
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121
pip install torch_geometric
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu121.html
pip install munch ruamel_yaml cilog gdown dive_into_graphs tensorboard rich typed-argument-parser

Usage

Supervised Training

You can use this command to run GCN on GOODArxiv dataset under supervised setting.

CUDA_VISIBLE_DEVICES=0 python supervised.py --config_path GOODCBAS/color/concept/ERM.yaml

Unsupervised Training

# baselines
python unsupervised.py --config_path GOODCBAS/color/concept/DGI.yaml

# Ours
python unsupervised.py --config_path GOODCBAS/color/concept/MARIO.yaml --ad_aug

python unsupervised.py --config_path GOODWebKB/university/concept/MARIO.yaml --ad_aug

Command Line Arguments

config_path is an important command line argument which specify the dataset, OOD type and learning algorithm. You can specify it as:

--config_path={dataset}/{domain}/{shift_type}/{method}.yaml

These options are available now.

dataset: GOODCora, GOODTwitch, GOODCBAS, GOODWebKB
domain: according to the selected dataset(GOODCora: word or degree, GOODTwitch: language)
shift_type: covariate, concept, no_shift
method: ERM, EERM, IRM, GAE, VGAE, GraphMAE, DGI, MVGRL, GRACE, BGRL, COSTA, MARIO

Hyperparameters

In the corresponding config yaml file (e.g., GOODCora/word/covariate/GRACE.yaml), you can assign training epochs (i.e., max_epoch) and learning rate (i.e., lr). And, for some methods, you can specify some additional hyperparametes (e.g., the strength of augmentation in GRACE and BGRL, temperature tau in GRACE, momentum mm and warmup steps in BGRL and etc).