y-sui / gslb

A comprehensive benchmark of Graph Structure Learning (NeurIPS 2023 Datasets and Benchmarks Track)

Graph Structure Learning Benchmark (GSLB)

GSLB is a Graph Structure Learning (GSL) library and benchmark based on DGL and PyTorch. We integrate diverse datasets and state-of-the-art GSL models.

📔 What is Graph Structure Learning?

Graph Structure Learning (GSL) aims to optimize both the parameters of Graph Neural Networks (GNNs) and the computation graph structure simultaneously. GSL methods start with input features and an optimal initial graph structure. Its corresponding computation graph is iteratively refined through a structure learning module. With the refined computation graph ,GNNs are used to generate graph representations. Then parameters of the GNNs and the structure modeling module are jointly updated, either simultaneously or alternatively, util a preset stopping condition is satisfied.

If you want to explore more information about GSL, please refer to our survey and paper collection.

🚀 Get Start

Requirements

GSLB needs the following requirements to be satisfied beforehand:

• Python 3.8+
• PyTorch 1.13
• DGL 1.1+
• Scipy 1.9+
• Scikit-learn
• Numpy
• NetworkX
• ogb
• tqdm
• easydict
• PyYAML
• DeepRobust

Installation via PyPI

To install GSLB with pip, simply run:

pip install GSLB

Then, you can import GSL from your current environment.

Usage

If you want to quickly run an existing GSL model on a graph dataset:

python main.py --dataset dataset_name --model model_name -n num_trails -g 0 --use_knn -k 5 --use_mettack --sparse --metric acc -p 0. --drop_rate 0. --add_rate 0.

Optional arguments:

--dataset : the name of graph dataset

--model : the name of GSL model

--ntrail -n : repetition count of experiments

--use_knn : whether to use knn graph instead of the original graph

--k : the number of the nearest neighbors

--drop_rate : the probability of randomly edge deletion

--add_rate : the probability of randomly edge addition

--mask_feat_rate : the probability of randomly mask features

--use_mettack : whether to use the structure after being attacked by mettack

--ptb_rate -p : the perturbation rate

--metric : the evaluation metric

--gpu_num -g : the selected GPU number

Example: Train GRCN on Cora dataset, with the evaluation metric is accuracy.

python main.py --dataset cora --model GRCN --metric acc

If you want to quickly generate a perturbed graph by Mettack:

python generate_attack.py --dataset cora --ptb_rate 0.05

Step 1: Load datasets

from GSL.data import *

# load a homophilic or heterophilic graph dataset
data = Dataset(root='/tmp/', name='cora')

# load a perturbed graph dataset
data = Dataset(root='/tmp/', name='cora', use_mettack=True, ptb_rate=0.05)

# load a heterogeneous graph dataset
data = HeteroDataset(root='/tmp/', name='acm')

# load a graph-level dataset
data = GraphDataset(root='/tmp/', name='IMDB-BINARY', model='GCN')

Step 2: Initialize the GSL model

from GSL.model import *
from GSL.utils import accuracy, macro_f1, micro_f1

model_name = 'GRCN'
metric = 'acc'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# the hyper-parameters are recorded in config
config_path = './configs/{}_config.yaml'.format(model_name.lower())

# select a evaluation metric
eval_metric = {
  'acc': accuracy,
  'macro-f1': macro_f1,
  'micro-f1': micro_f1
}[metric]

model = GRCN(data.num_feat, data.num_class, evel_metric,
            config_path, dataset_name, device)

Step 3: Train GSL model

model.fit(data)

🧩 Implementation Algorithms

Currently, we have implemented the following GSL algorithms:

Algorithm	Conference/Journal	Category	Paper	Code
LDS	ICML 2019	Homogeneous GSL	Learning Discrete Structures for Graph Neural Networks	Link
GRCN	ECML PKDD 2020	Homogeneous GSL	Graph-Revised Convolutional Network	Link
ProGNN	KDD 202	Homogeneous GSL	Graph structure learning for robust graph neural networks	Link
IDGL	NeurIPS 2020	Homogeneous GSL	Iterative Deep Graph Learning for Graph Neural Networks: Better and Robust Node Embeddings	Link
GEN	WWW 2021	Homogeneous GSL	Graph Structure Estimation Neural Networks	Link
CoGSL	WWW 2022	Homogeneous GSL	Compact Graph Structure Learning via Mutual Information Compression	Link
SLAPS	NeurIPS 2021	Homogeneous GSL	SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks	Link
SUBLIME	WWW 2022	Homogeneous GSL	Towards Unsupervised Deep Graph Structure Learning	Link
STABLE	KDD 2022	Homogeneous GSL	Reliable Representations Make A Stronger Defender: Unsupervised Structure Refinement for Robust GNN	Link
NodeFormer	NeurIPS 2022	Homogeneous GSL	NodeFormer: A Scalable Graph Structure Learning Transformer for Node Classification	Link
HES-GSL	TNNLS 2023	Homogeneous GSL	Homophily-Enhanced Self-Supervision for Graph Structure Learning: Insights and Directions	Link
GSR	WSDM 2023	Homogeneous GSL	Self-Supervised Graph Structure Refinement for Graph Neural Networks	Link
GTN	NeurIPS 2020	Heterogeneous GSL	Graph Transformer Networks	Link
HGSL	AAAI 2021	Heterogeneous GSL	Heterogeneous Graph Structure Learning for Graph Neural Networks	Link
HGP-SL	AAAI 2020	Graph-level GSL	Hierarchical Graph Pooling with Structure Learning	Link
VIB-GSL	AAAI 2022	Graph-level GSL	Graph Structure Learning with Variational Information Bottleneck	Link