Hierarchical Prototype Networks for Continual Graph Representation Learning (HPNs)

Implementations of Hierarchical Prototype Networks for Continual Graph Representation Learning. paper

@article{Zhang23TPAMI,
 title        = {Hierarchical Prototype Networks for Continual Graph Representation Learning},
 author       = {Xikun Zhang and Dongjin Song and Dacheng Tao},
 journal      = {{IEEE} Trans. Pattern Anal. Mach. Intell.},
 volume       = {45},
 number       = {4},
 pages        = {4622--4636},
 year         = {2023},
 url          = {https://doi.org/10.1109/TPAMI.2022.3186909},
 doi          = {10.1109/TPAMI.2022.3186909},
}

Dependencies

Our final test was done with the following configurations. However, early versions of these packages may also work.

• Numpy == 1.19.1
• PyTorch == 1.7.1
• ogb == 1.3.1

Data preprocessing

All datasets are uploaded except OGB-Arxiv and OGB-Products which are too large for github. For these two datasets, the preprocessing codes below provide the entire process for downloading and preprocessing (splitting datasets into a sequence of tasks). On these large datasets, to avoid wasting time on preprocessing each time running the programs, the code will also store the preprocessed data. If you wish to preprocess both OGB-Arxiv and OGB-Products together, please run the following command.

bash OGB_preprocess.sh

Or if you want to preprocess one of them, please run the following commands.

python OGB_preprocess.py --data_name ogbn-arxiv

python OGB_preprocess.py --data_name ogbn-products

Run experiments

run experiments on Cora

bash run_cora.sh

run experiments on Citeseer

bash run_citeseer.sh

run experiments on Actor

bash run_actor.sh

run experiments on OGB-Arxiv

bash run_arxiv.sh

run experiments on OGB-Products

bash run_products.sh