This repository is the implementation of CTGC, a self-supervised graph condensation approach designed to efficiently handle diverse downstream tasks.

All experiments are implemented in Python 3.9 with Pytorch 1.11.0.

To install the required dependencies, run:

pip install -r requirements.txt

Following is the step-by-step instruction for our proposed method CTGC.

Set the dataset folder path using the --data_dir (defaut: ./data/).

• For Cora and CiteSeer, the code will directly download from PyG.
• For Ogbn-arxiv and Reddit, we use the datasets provided by GraphSAINT, which are available on Google Drive or BaiduYun (code: f1ao).

Prepare the dataset and perform eigenvalue decomposition by running:

$ bash ./scr/run_preprocess.sh

• Split data files will be saved in the folder ./dataset_split/.
• Eigenvectors and eigenvalues will be saved in the folder ./save_eigen/.

Pre-train the semantic relay model using the following command:

$ bash ./scr/run_pretrain.sh

The pre-trained relay model will be saved in the folder: ./save_pretrain_model/.

Iteratively train the semantic and structural relay models by executing:

$ bash ./scr/run_relay.sh

The trained relay models and centroid embeddings will be stored in the directory: ./save_pretrain_model/.

Generate the condensed graph by running:

$ bash ./scr/run_condense.sh

The generated graph will be saved in the folder ./save_condensed_data/.

The performances for the pre-trained relay model, semantic relay model, and the condensed graph (from Steps 3–5) are recorded in CSV files located in the folder: ./results_proposed/.

To evaluate the models trained on the original graph, run:

$ bash ./scr/run_whole.sh

The performances are recorded in the folder: ./results_whole/.

@inproceedings{gao2024contrastive,
  title={Contrastive graph condensation: Advancing data versatility through self-supervised learning},
  author={Gao, Xinyi and Li, Yayong and Chen, Tong and Ye, Guanhua and Zhang, Wentao and Yin, Hongzhi},
  booktitle={Proceedings of the 31th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year={2025}
}