GraphMAE: Self-Supervised Masked Graph Autoencoders

Implementation for KDD'22 paper: GraphMAE: Self-Supervised Masked Graph Autoencoders.

We also have a Chinese blog about GraphMAE on Zhihu (知乎), and an English Blog on Medium.

GraphMAE is a generative self-supervised graph learning method, which achieves competitive or better performance than existing contrastive methods on tasks including node classification, graph classification, and molecular property prediction.

❗ Update

[2023-04-12] GraphMAE2 is published and the code can be found here.

[2022-12-14] The PYG implementation of GraphMAE for node / graph classification is available at this branch.

Dependencies

Python >= 3.7
Pytorch >= 1.9.0
dgl >= 0.7.2
pyyaml == 5.4.1

Quick Start

For quick start, you could run the scripts:

Node classification

sh scripts/run_transductive.sh <dataset_name> <gpu_id> # for transductive node classification
# example: sh scripts/run_transductive.sh cora/citeseer/pubmed/ogbn-arxiv 0
sh scripts/run_inductive.sh <dataset_name> <gpu_id> # for inductive node classification
# example: sh scripts/run_inductive.sh reddit/ppi 0

# Or you could run the code manually:
# for transductive node classification
python main_transductive.py --dataset cora --encoder gat --decoder gat --seed 0 --device 0
# for inductive node classification
python main_inductive.py --dataset ppi --encoder gat --decoder gat --seed 0 --device 0

Supported datasets:

transductive node classification: cora, citeseer, pubmed, ogbn-arxiv
inductive node classification: ppi, reddit

Run the scripts provided or add --use_cfg in command to reproduce the reported results.

Graph classification

sh scripts/run_graph.sh <dataset_name> <gpu_id>
# example: sh scripts/run_graph.sh mutag/imdb-b/imdb-m/proteins/... 0 

# Or you could run the code manually:
python main_graph.py --dataset IMDB-BINARY --encoder gin --decoder gin --seed 0 --device 0

Supported datasets:

IMDB-BINARY, IMDB-MULTI, PROTEINS, MUTAG, NCI1, REDDIT-BINERY, COLLAB

Run the scripts provided or add --use_cfg in command to reproduce the reported results.

Molecular Property Prediction

Please refer to codes in ./chem for molecular property prediction.

Datasets

Datasets used in node classification and graph classification will be downloaded automatically from https://www.dgl.ai/ when running the code.

Experimental Results

Node classification (Micro-F1, %):

	Cora	Citeseer	PubMed	Ogbn-arxiv	PPI	Reddit
DGI	82.3±0.6	71.8±0.7	76.8±0.6	70.34±0.16	63.80±0.20	94.0±0.10
MVGRL	83.5±0.4	73.3±0.5	80.1±0.7	-	-	-
BGRL	82.7±0.6	71.1±0.8	79.6±0.5	71.64±0.12	73.63±0.16	94.22±0.03
CCA-SSG	84.0±0.4	73.1±0.3	81.0±0.4	71.24±0.20	73.34±0.17	95.07±0.02
GraphMAE(ours)	84.2±0.4	73.4±0.4	81.1±0.4	71.75±0.17	74.50±0.29	96.01±0.08

Graph classification (Accuracy, %)

	IMDB-B	IMDB-M	PROTEINS	COLLAB	MUTAG	REDDIT-B	NCI1
InfoGraph	73.03±0.87	49.69±0.53	74.44±0.31	70.65±1.13	89.01±1.13	82.50±1.42	76.20±1.06
GraphCL	71.14±0.44	48.58±0.67	74.39±0.45	71.36±1.15	86.80±1.34	89.53±0.84	77.87±0.41
MVGRL	74.20±0.70	51.20±0.50	-	-	89.70±1.10	84.50±0.60	-
GraphMAE(ours)	75.52±0.66	51.63±0.52	75.30±0.39	80.32±0.46	88.19±1.26	88.01±0.19	80.40±0.30

Transfer learning on molecular property prediction (ROC-AUC, %):

	BBBP	Tox21	ToxCast	SIDER	ClinTox	MUV	HIV	BACE	Avg.
AttrMasking	64.3±2.8	76.7±0.4	64.2±0.5	61.0±0.7	71.8±4.1	74.7±1.4	77.2±1.1	79.3±1.6	71.1
GraphCL	69.7±0.7	73.9±0.7	62.4±0.6	60.5±0.9	76.0±2.7	69.8±2.7	78.5±1.2	75.4±1.4	70.8
GraphLoG	72.5±0.8	75.7±0.5	63.5±0.7	61.2±1.1	76.7±3.3	76.0±1.1	77.8±0.8	83.5±1.2	73.4
GraphMAE(ours)	72.0±0.6	75.5±0.6	64.1±0.3	60.3±1.1	82.3±1.2	76.3±2.4	77.2±1.0	83.1±0.9	73.8

Citing

If you find this work is helpful to your research, please consider citing our paper:

@inproceedings{hou2022graphmae,
  title={GraphMAE: Self-Supervised Masked Graph Autoencoders},
  author={Hou, Zhenyu and Liu, Xiao and Cen, Yukuo and Dong, Yuxiao and Yang, Hongxia and Wang, Chunjie and Tang, Jie},
  booktitle={Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={594--604},
  year={2022}
}