arXiv (arXiv:2406.142886),
- (Modularity maximization == contrastive learning) We establish the connection between modularity maximization and graph contrastive learning
- (MAGI framework) We propose MAGI, a community-aware graph contrastive learning framework that uses modularity maximization as its pretext task
- (Performance and scalibility) MAGI has achieved state-of-the-art performance in graph clustering and demonstrates excellent scalability on industrial-scale graphs
Note
Higher versions should be also compatible.
- PyTorch
- PyTorch Geometric
- PyTorch Cluster
- PyTorch Scatter
- PyTorch Sparse
- Scipy
- Scikit-learn
- Scikit-learn-intelex
- Ogb
pip install -r requirements.txt
- Cora
python train_gcn.py --runs 10 --dataset 'Cora' --hidden '512' --wt 100 --wl 2 --tau 0.3 --ns 0.5 --lr 0.0005 --epochs 400 --wd 1e-3
- CiteSeer
python train_gcn.py --runs 10 --dataset 'Citeseer' --hidden '1024,512' --wt 100 --wl 3 --tau 0.9 --ns 0.5 --lr 0.0001 --epochs 400 --wd 5e-4
- Amazon-photo
python train_gcn.py --runs 10 --dataset 'Photo' --hidden '512' --wt 100 --wl 3 --tau 0.5 --ns 0.5 --lr 0.0005 --epochs 400 --wd 1e-3
- Amazon-computers
python train_gcn.py --runs 10 --dataset 'Computers' --hidden '1024,512' --wt 100 --wl 3 --tau 0.9 --ns 0.1 --lr 0.0005 --epochs 400 --wd 1e-3
- ogbn-arxiv
python train_sage.py --runs 10 --dataset 'ogbn-arxiv' --batchsize 2048 --max_duration 60 --kmeans_device 'cpu' --kmeans_batch -1 --hidden '1024,256' --size '10,10' --wt 20 --wl 5 --tau 0.9 --ns 0.1 --lr 0.01 --epochs 400 --wd 0 --dropout 0
- reddit
python train_sage.py --runs 10 --dataset 'Reddit' --batchsize 2048 --max_duration 60 --kmeans_device 'cpu' --kmeans_batch -1 --hidden '1024,256' --size '10,10' --wt 20 --wl 5 --tau 0.5 --ns 0.5 --lr 0.01 --epochs 400 --wd 0 --dropout 0
- ogbn-products
python train_sage.py --runs 10 --dataset 'ogbn-products' --batchsize 2048 --max_duration 60 --kmeans_device 'cuda' --kmeans_batch 300000 --hidden '1024,1024,256' --size '10,10,10' --wt 20 --wl 4 --tau 0.9 --ns 0.1 --lr 0.01 --epochs 400 --wd 0 --dropout 0
Please kindly cite our paper if you find our project is useful:
@inproceedings{magi,
title = {Revisiting Modularity Maximization for Graph Clustering: A Contrastive Learning Perspective},
author = {Yunfei Liu and Jintang Li and Yuehe Chen and Ruofan Wu and Baokun Wang and Jing Zhou and Sheng Tian and Shuheng Shen and Xing Fu and Changhua Meng and Weiqiang Wang and Liang Chen},
booktitle = {{KDD}},
publisher = {{ACM}},
year = {2024},
}