This is the source code for paper ''A Topology-aware Graph Coarsening Framework for Continual Graph Learning'' to appear in The Thirty-eighth Annual Conference on Neural Information Processing Systems (Neurips 2024).

Xiaoxue Han, Zhuo Feng, Yue Ning

The code has been successfully tested in the following environment. (For older versions, you may need to modify the code)

• Python 3.8.13
• PyTorch 1.12.1+cu11.6
• dgl 0.9.1
• pygsp 0.5.1
• sklearn 1.1.2

• Kindle dataset: https://jmcauley.ucsd.edu/data/amazon/amazon_readme.txt
• DBLP dataset: https://originalstatic.aminer.cn/misc/dblp.v13.7z
• ACM dataset: https://lfs.aminer.org/lab-datasets/citation/acm.v9.zip

Place the download files to raw-data folder. The folder structure is as follows:

- TACO-code
	- raw-data
		- data file
		- ...
	- src
  - ...

To preprocess the data for each dataset, please run the following commands in order under the [dataset-name]-data-preprocessing folder:

python 0-process_dataset.py (only for ACM and DBLP datasets)

python 1-build_graph.py

python 2-get_largest_connected_component.py

python 3-generate_random_masks.py

Please run following commands for training and testing under the src folder. We take the dataset kindle with GCN as backbone GNN model as the example.

Evaluate the TACO model

python -W ignore train.py --dataset kindle --method DYGRA --gnn GCN --reduction_rate 0.5 --buffer_size 200

Please cite our paper if you find this code useful for your research:

BibTeX

@misc{han2024topologyawaregraphcoarseningframework,
      title={A Topology-aware Graph Coarsening Framework for Continual Graph Learning}, 
      author={Xiaoxue Han and Zhuo Feng and Yue Ning},
      year={2024},
      eprint={2401.03077},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2401.03077}, 
}