egc 0.3.0

Benchmark of Graph Clustering.

EAGLEGraphClustering

A benchmark of graph clustering from EAGLE-Lab, Zhejiang University.

• Installation
• Usage
  • Pip Package
  • Command line
• Datasets
• Implemented baseline methods
  • Disjoint
    • Unsupervised Graph Neural Networks + Kmeans
    • End-to-End Graph Clustering
  • Overlapping
• Requirements
• Contributing

Installation

• python>=3.8
• torch>=1.12
• dgl>=1.1

$ python -m pip install egc

Usage

Pip Package

See egc for docs.

• Import the package and use any graph clustering model supported as:

from torch import nn
from egc.model import DGL_GAEKmeans
from egc.utils import load_data
from egc.utils import get_default_args
from egc.utils import set_seed
from egc.utils import set_device

# set the random seed
set_seed(4096)

# set the gpu id
set_device('0')

# load graph
graph, label, n_clusters = load_data(
    dataset_name='Cora',
    directory='./data',
)
features = graph.ndata["feat"]
adj_csr = graph.adj_external(scipy_fmt="csr")

# get default args
args = get_default_args('gae_kmeans')

# init the model
model = DGL_GAEKmeans(
    epochs=args["epochs"],
    n_clusters=10,
    fead_dim=features.shape[1],
    n_nodes=features.shape[0],
    hidden_dim1=args["hidden1"],
    dropout=args["dropout"],
    lr=args["lr"],
    early_stop=args["early_stopping_epoch"],
    activation=args["activation"],
)

# fit the model
model.fit(adj_csr, features)

# get clustering results
res = model.get_memberships()

Command line

• Clone the Repo
• Install the env

# NOTE: python>=3.8 is needed
# Install cuda if necessary. Check Cuda version first.
$ cd EGC
# Leave out `bash .ci/install-dev.sh &&` if no dev env is needed.
$ bash .ci/install-dev.sh && bash .ci/install.sh
# run `source .env/bin/activate` to activate the virtual env

• Run any supported model as:

$ python train.py ${OPTIONAL global args} ${POSITIONAL args (model)} ${optional model args}

• OPTIONAL global args which should be used before ${model}
• POSITIONAL args, i.e., all models supported

E.g.,

# check OPTIONAL global args, e.g., all models supported
$ python train.py -h

# check optional args of certain model
$ python train.py gae_kmeans -h

# run any model
$ python train.py --dataset=Cora gae_kmeans --lr 0.001

Datasets

• Cora, Citeseer, Pubmed come from DGL Lib
• BlogCatalog, Flickr come from CoLA github
• ACM come from SDCN github
• All above datasets are converted to undirected graphs.

Dataset	Nodes	Edges	Attributes	Classes
Cora	2,708	10,556	1,433	7
Citeseer	3,327	9,228	3,703	6
Pubmed	19,717	88,651	500	3
BlogCatalog	5,196	343,486	8,189	6
Flickr	7,575	479,476	12,047	9
ACM	3,025	26,256	1,870	3
CoraFull	19,793	126,842	8,710	70

Implemented baseline methods

Disjoint

Unsupervised Graph Neural Networks + Kmeans

method	Conf/Journal	Original Code	Supproted
VGAE	16nips	TensorFlow	✅
GraphSAGE
DGI	19iclr	Pytorch	✅
GMI	20www	Pytorch	✅
SENet	21nn		✅

End-to-End Graph Clustering

method	Conf/Journal	Original Code	Supproted
SDCN	20www	Pytorch	✅
DANMF	18cikm	code	✅
M-NMF	17aaai	Matlab	✅
DFCN	21aaai	Pytorch	✅
VGAECD	18icdm	--	✅
ComE	17cikm	code	✅
DAEGC	19ijcai	code	✅

Overlapping

method	Conf/Journal	Original Code	Supproted
CommunityGAN	19www	TensorFlow	✅

Requirements

See dependencies, requirements-dev.txt and requirements.txt.

Contributing

See CONTRIBUTING.md.