Graph Data Condensation via Self-expressive Graph Structure Reconstruction (KDD 2024)

[KDD 2024] In this repository, we present the code of Graph Data Condensation via Self-expressive Graph Structure Reconstruction (GCSR).

Data

For cora and citeseer, the code will directly download them using PyG (PyTorch Geometric). For reddit, flickr and ogbn-arxiv, we use the datasets provided by GraphSAINT. They are available on Google Drive link (reddit, flickr and ogbn-arxiv). Download and move them to ./data at your root directory.

Environment

python==3.7.13
torch==1.13.0
torchvision==0.14.0
torch_geometric==2.3.1
scikit_learn==1.0.2
scipy==1.7.3
numpy==1.21.6
ogb==1.3.6
deeprobust==0.2.8
torch_sparse==0.6.17
torch_scatter==2.1.1

Reproducibility

The generated graphs are saved in the folder ./saved_ours. You can directly load them to test the performance.

• For Table 1, run bash ./script/main_table.sh.

It should be noted that, for each dataset and each condensation architecture, the training trajectory buffer should only be produced one time. E.g., for citeseer condensed by SGC, you just need to run the following command one time:

python buffer.py --dataset=citeseer --model=sgc2-lr3-wt54 --model_name=SGC2 --num_experts=100 --lr_teacher=1e-3

If you want to test the synthetic graph while condensing, add --test to each condensation command, e.g.:

python condense.py --test --gpu_id=0 --dataset=citeseer --expert_net=sgc2-lr3-wt54 --expert_net_type=SGC2 --test_net_type=GCN --epochs=1000 --eval_interval=200 --student_epochs=5 --max_start_epoch=60 --expert_epochs=2 --lr_feat=1e-6 --reduction_rate=0.5 --saved_folder=saved_ours --normalize --with_val --exps=5 --alpha=1 --beta=0.999 --tau=0.9 --gamma=0.5 --message_passing=4 --dropout_test=0.0

and add --save if you want to save your synthetic graphs. student_epochs denotes N in the paper and expert_epochs denotes M in the paper. Both of them are hyperparameters for multi-step gradient matching.

• For Table 2, run bash ./script/dif_test.sh.

• For Table 3, run bash ./script/dif_condense.sh.

• For ablation study and Table 4, run bash ./script/ablation.sh.

Citation

If you find this repo useful in your research, please consider citing our paper as follows:

@article{liu2024graph,
  title={Graph Data Condensation via Self-expressive Graph Structure Reconstruction},
  author={Liu, Zhanyu and Zeng, Chaolv and Zheng, Guanjie},
  journal={arXiv preprint arXiv:2403.07294},
  year={2024}
}