Modularity-Aware Graph Autoencoders

This repository provides a Tensorflow implementation of the four models:

• Modularity-aware graph autoencoder with linear encoder;
• Modularity-aware graph autoencoder with 2-layer GCN encoder;
• Modularity-aware variational graph autoencoder with linear encoder;
• Modularity-aware variational graph autoencoder with 2-layer GCN encoder;

introduced in the article Modularity-Aware Graph Autoencoders for Joint Community Detection and Link Prediction. This article has been accepted for publication in Elsevier's Neural Networks journal in 2022. It has also been presented at the NeurIPS 2022 Workshop on New Frontiers in Graph Learning (GLFrontiers) in New Orleans.

We provide the Blogs, Cora, Cora-Large, Citeseer, Pubmed, and SBM datasets in the data folder. The seventh dataset from our experiments, i.e., Deezer-Album, is a private dataset that we do not release in this repository.

This code extensively builds upon the linear GAE/VGAE and FastGAE repositories by Deezer, themselves building upon the original Tensorflow implementation of GAE and VGAE by Thomas Kipf.

Installation

git clone https://github.com/GuillaumeSalhaGalvan/modularity_aware_gae
cd modularity_aware_gae
python setup.py install

Requirements: networkx, numpy, python-louvain, scikit-learn, scipy, tensorflow (1.X).

Run Experiments

A first quick test on Citeseer (20 training iterations only):

cd modularity_aware_gae
python train.py --dataset=citeseer --features=False --task=task_2 --model=linear_vae --iterations=20 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=0.5 --lamb=0.75 --gamma=0.5 --s_reg=2 --fastgae=False --nb_run=1

Modularity-Aware models with settings from Table 7:

cd modularity_aware_gae
python train.py --dataset=blogs --features=False --task=task_2 --model=gcn_vae --iterations=200 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=0.75 --lamb=0.5 --gamma=2 --s_reg=10 --fastgae=False --nb_run=1
python train.py --dataset=citeseer --features=False --task=task_2 --model=linear_vae --iterations=500 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=0.5 --lamb=0.75 --gamma=0.5 --s_reg=2 --fastgae=False --nb_run=1
python train.py --dataset=citeseer --features=True --task=task_2 --model=linear_vae --iterations=500 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=0.5 --lamb=0.75 --gamma=0.5 --s_reg=2 --fastgae=False --nb_run=1
python train.py --dataset=pubmed --features=False --task=task_2 --model=linear_ae --iterations=500 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=0.5 --lamb=0.1 --gamma=0.1 --s_reg=5 --fastgae=False --nb_run=1
python train.py --dataset=pubmed --features=True --task=task_2 --model=linear_vae --iterations=700 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=0.5 --lamb=0.1 --gamma=10 --s_reg=2 --fastgae=False --nb_run=1
python train.py --dataset=cora --features=False --task=task_2 --model=linear_ae --iterations=500 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=1.0 --lamb=0.25 --gamma=0.25 --s_reg=1 --fastgae=False --nb_run=1
python train.py --dataset=cora --features=True --task=task_2 --model=linear_vae --iterations=300 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=0.01 --lamb=0.001 --gamma=1 --s_reg=1 --fastgae=False --nb_run=1
python train.py --dataset=cora-large --features=False --task=task_2 --model=linear_vae --iterations=500 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=0.1 --lamb=0.001 --gamma=0.1 --s_reg=10 --fastgae=False --nb_run=1
python train.py --dataset=sbm --features=False --task=task_2 --model=linear_vae --iterations=300 --learning_rate=0.01 --hidden=32 --dimension=16 --beta=0.1 --lamb=0.5 --gamma=2 --s_reg=10 --fastgae=True --nb_node_samples=10000 --measure=degree --alpha=1.0 --nb_run=1

Notes:

• The above commands will complete a single training for each model. Use the nb_run parameter to average scores over several model trainings.
• The above commands will evaluate results on "Task 2" from the paper, i.e., joint community detection and link prediction. Set --task=task_1 to report results on "Task 1", i.e., pure community detection.
• The complete list of parameters is available in the train.py file.

Cite

Please cite our paper if you use this code in your own work:

@article{salhagalvan2022modularity,
  title={Modularity-Aware Graph Autoencoders for Joint Community Detection and Link Prediction},
  author={Salha-Galvan, Guillaume and Lutzeyer, Johannes F and Dasoulas, George and Hennequin, Romain and Vazirgiannis, Michalis},
  journal={Neural Networks},
  volume={153},
  pages={474--495},
  year={2022},
  publisher={Elsevier}
}