Overview • Installation • Examples • Docs • Citation

OpenGSL

Official code for OpenGSL: A Comprehensive Benchmark for Graph Structure Learning. OpenGSL is a comprehensive benchmark for Graph Structure Learning(GSL). GSL is a family of data-centric learning approaches which jointly optimize the graph structure and the corresponding GNN models. It has great potential in many real-world applications, including disease analysis, protein structure prediction, etc.

Overview of the Benchmark

OpenGSL provides a fair and comprehensive platform to evaluate existing GSL works and facilitate future GSL research.

Installation

Note: OpenGSL depends on PyTorch, PyTorch Geometric, PyTorch Sparse and DEEP GRAPH LIBRARY (DGL). To streamline the installation, OpenGSL does NOT install these libraries for you. Please install them from the above links for running OpenGSL:

torch>=1.9.1
torch_geometric>=2.1.0
torch_sparse>=0.6.12
dgl>=0.9.0

Using Pip

pip install opengsl

Installation for local development:

git clone https://github.com/OpenGSL/OpenGSL
cd opengsl
pip install -e .

Required Dependencies:

Python 3.7+
ruamel.yaml
pandas
scipy
scikit-learn
pyro-api
pyro-ppl
numba

Quick Start

You can use the command python examples/gcn_cora.py or follow the 4 steps.

The following example shows you how to perform GCN on the Cora dataset.

Step 1: Load configuration

import opengsl
conf = opengsl.config.load_conf(method="gcn", dataset="cora")

Method and Dataset parameters in Built-in configuration

method ： gcn, prognn, idgl, grcn, gaug, slaps, gt, nodeformer, gen, cogsl, segsl, sublime, stable

dataset ： cora, pubmed, citeseer, blogcatalog, flickr, amazon-ratings, questions, minesweeper, roman-empire, wiki-cooc

Step 2: Load data

dataset = opengsl.data.Dataset("cora", n_splits=1, feat_norm=conf.dataset['feat_norm'])

Step 3: Build Model

solver = opengsl.method.GCNSolver(conf,dataset)

Step 4: Training and Evaluation

exp = opengsl.ExpManager(solver)
exp.run(n_runs = 10)

Adding New GSL Method

if you want to use your own GSL method, see customized_gsl.py for detail.

Node Classification Results

Model	Cora	Citeseer	Pubmed	Questions	Minesweeper
GCN	81.95 ± 0.62	71.34 ± 0.48	78.98 ± 0.35	75.80 ± 0.51	78.28 ± 0.44
ProGNN	80.27 ± 0.48	71.35 ± 0.42	79.39 ± 0.29	--	51.43 ± 2.22
IDGL	84.19 ± 0.61	73.26 ± 0.53	82.78 ± 0.44	50.00 ± 0.00	50.00 ± 0.00
GRCN	84.61 ± 0.34	72.34 ± 0.73	79.30 ± 0.34	74.50 ± 0.84	72.57 ± 0.49
GAug(O)	83.43 ± 0.53	72.79 ± 0.86	78.73 ± 0.77	--	77.93 ± 0.64
SLAPS	72.29 ± 1.01	70.00 ± 1.29	70.96 ± 0.99	--	50.89 ± 1.72
GT	80.79 ± 1.14	68.50 ± 2.07	77.91 ± 0.58	75.36 ± 0.92	89.70 ± 0.28
Nodeformer	78.81 ± 1.21	70.39 ± 2.04	78.38 ± 1.94	72.61 ± 2.29	77.29 ± 1.71
GEN	81.66 ± 0.91	73.21 ± 0.62	78.49 ± 3.98	--	79.56 ± 1.09
CoGSL	81.46 ± 0.88	72.94 ± 0.71	--	--	--
SEGSL	81.04 ± 1.07	71.57 ± 0.40	79.26 ± 0.67	--	--
SUBLIME	83.33 ± 0.73	72.44 ± 0.89	80.56 ± 1.32	67.21 ± 0.99	49.93 ± 1.36
STABLE	83.25 ± 0.86	70.99 ± 1.19	81.46 ± 0.78	--	70.78 ± 0.27

Model	BlogCatalog	Flickr	Amazon-ratings	Roman-empire	Wiki-cooc
GCN	76.12 ± 0.42	61.60 ± 0.49	45.24 ± 0.29	70.41 ± 0.47	92.03 ± 0.19
ProGNN	73.38 ± 0.30	52.88 ± 0.76	--	56.21 ± 0.58	89.07 ± 5.59
IDGL	89.68 ± 0.24	86.03 ± 0.25	45.87 ± 0.58	47.10 ± 0.65	90.18 ± 0.27
GRCN	76.08 ± 0.27	59.31 ± 0.46	50.06 ± 0.38	44.41 ± 0.41	90.59 ± 0.37
GAug(O)	76.92 ± 0.34	61.98 ± 0.67	48.42 ± 0.39	52.74 ± 0.48	91.30 ± 0.23
SLAPS	91.73 ± 0.40	83.92 ± 0.63	40.97 ± 0.45	65.35 ± 0.45	89.09 ± 0.54
GT	70.70 ± 5.62	43.19 ± 6.53	48.55 ± 0.34	76.49 ± 0.80	90.26 ± 1.24
Nodeformer	44.53 ± 22.62	67.14 ± 6.77	41.33 ± 1.25	56.54 ± 3.73	54.83 ± 4.43
GEN	90.48 ± 0.99	84.84 ± 0.81	49.17 ± 0.68	--	91.15 ± 0.49
CoGSL	--	--	--	--	--
SEGSL	75.03 ± 0.28	60.59 ± 0.54	--	--	--
SUBLIME	95.29 ± 0.26	88.74 ± 0.29	44.49 ± 0.30	63.93 ± 0.27	76.10 ± 1.12
STABLE	71.84 ± 0.56	51.36 ± 1.24	48.36 ± 0.21	41.00 ± 1.18	80.46 ± 2.44

How to Contribute

As an active research topic, we are witnessing the rapid development of GSL methods. Hence, this project will be frequently updated, and we welcome everyone interested in this topic to contribute!

Please feel free to send PR or issue!

Citation

Our paper on this benchmark will be released soon!

If you use our benchmark in your works, we would appreciate citations to the paper:

@article{zhou2023opengsl,
  title={OpenGSL: A Comprehensive Benchmark for Graph Structure Learning},
  author={Zhiyao Zhou, Sheng Zhou, Bochao Mao, Xuanyi Zhou, Jiawei Chen, Qiaoyu Tan, Daochen Zha, Can Wang, Yan Feng, Chun Chen},
  journal={arXiv preprint arXiv:2306.10280},
  year={2023}
}

Reference

ID	Paper	Method	Conference
1	Semi-supervised classification with graph convolutional networks	GCN	ICLR 2017
2	Graph Structure Learning for Robust Graph Neural Networks	ProGNN	KDD 2020
3	Iterative Deep Graph Learning for Graph Neural Networks: Better and Robust Node Embeddings	IDGL	NeurIPS 2020
4	Graph-Revised Convolutional Network	GRCN	ECML-PKDD 2020
5	Data Augmentation for Graph Neural Networks	GAug(O)	AAAI 2021
6	SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks	SLAPS	ICML 2021
7	Masked label prediction: Unified message passing model for semi-supervised classification	GT	IJCAI 2021
8	Nodeformer: A scalable graph structure learning transformer for node classification	Nodeformer	NeurIPS 2022
9	Graph Structure Estimation Neural Networks	GEN	WWW 2021
10	Compact Graph Structure Learning via Mutual Information Compression	CoGSL	WWW 2022
11	SE-GSL: A General and Effective Graph Structure Learning Framework through Structural Entropy Optimization	SEGSL	WWW 2023
12	Towards Unsupervised Deep Graph Structure Learning	SUBLIME	WWW 2022
13	Reliable Representations Make A Stronger Defender: Unsupervised Structure Refinement for Robust GNN	STABLE	KDD 2022

yorktown-class / opengsl Goto Github PK

opengsl's Introduction