Pytorch implementation of paper 'GraphSMOTE: Imbalanced Node Classification on Graphs with Graph Neural Networks' on WSDM2021
- python3
- ipdb
- pytorch1.0
- network 2.4
- scipy
- sklearn
Two processed datasets are published, including BlogCatalog and Cora. For downloading them, please click. Please unzip them in the 'data' folder.
The twitter dataset is obtained from link, with fake id. For twitter dataset, node embeddings can be generated with Deepwalk, following this repo. Loading the twitter dataset is also provided in data_load.py.
We provide two base architectures, GCN and GraphSage. The default one is GraphSage, and can be set via '--model'.
The default value is 1. If want to make every class balanced instead of using pre-set ratios, please set it to 0 in '--up_scale'.
During finetune, set '--setting='newG_cls'' correponds to use pretrained decoder, and set '--setting='recon_newG'' corresponds to also finetune the decoder.
Besides, during finetune, '--opt_new_G' corresponds to update decoder with also classification losses. This option may cause more variance in performance, and usually need more careful hyper-parameter choices.
Below is an example for the Cora dataset.
- Pretrain the auto-encoder
python main.py --imbalance --no-cuda --dataset=cora --setting='recon'
Pretrained model can be found in the corresponding checkpoint folder. Rename and set the path to pretrained checkpoint as [dataset]\Pretrained.pth
- Finetune
python main.py --imbalance --no-cuda --dataset=cora --setting='newG_cls' --load=Pretrained.pth
We provide four baselines in this code. They can be configured via the '--setting' arguments. Please refer to the 'get_parser()' function in utils.py.
- Oringinal model: Vanilla backbone models. '--setting='no''
- Over-sampling in raw input domain: Repeat nodes in the minority classes. '--setting='upsampling''
- Reweight: Give samples from minority classes a larger weight when calculating the loss. '--setting='reweight''
- Embed-SMOTE: Perform SMOTE in the intermediate embedding domain. '--setting='embed_up''
Use over-sampling as an example:
python main.py --imbalance --no-cuda --dataset=cora --setting='upsampling'
If any problems occur via running this code, please contact us at tkz5084@psu.edu.
Thank you!