lirongwu / graphmixup

Code for ECML-PKDD 2022 paper "GraphMixup: Improving Class-Imbalanced Node Classification by Reinforcement Mixup and Self-supervised Context Prediction"

GraphMixup

This is a PyTorch implementation of the GraphMixup, and the code includes the following modules:

• Dataset Loader (Cora, BlagCatalog, and Wiki-CS)

• Various Architectures (GCN, SAGE, GAT, and SEM)

• Five compared baselines (Origin, Over-Sampling, Re-weight, SMOTE, and Embed-SMOTE)

• Training paradigm (joint learning, pre-training, and fine-tuning) for node classification on three datasets

• Visualization and evaluation metrics

Main Requirements

• networkx==2.5
• numpy==1.19.2
• scikit-learn==0.24.1
• scipy==1.5.2
• torch==1.6.0

Description

• train.py

  • train() -- Train a new model for node classification task on the Cora, BlagCatalog, and Wiki-CS datasets
  • test() -- Test the learned model for node classification task on the Cora, BlagCatalog, and Wiki-CS datasets
  • save_model() -- Save the pre-trained model
  • load_model() -- Load model for fine-tuning

• data_load.py

  • load_cora() -- Load Cora Dataset
  • load_BlogCatalog() -- Load BlogCatalog Dataset
  • load_wiki_cs() -- Load Wiki-CS Dataset

• models.py

  • GraphConvolution() -- GCN Layer
  • SageConv() -- SAGE Layer
  • SemanticLayer() -- Semantic Feature Layer
  • GraphAttentionLayer() -- GAT Layer
  • PairwiseDistance() -- Perform self-supervised Local-Path Prediction
  • DistanceCluster() -- Perform self-supervised Global-Path Prediction

• utils.py

  • src_upsample() -- Perform interpolation in the input space
  • src_smote() -- Perform interpolation in the embedding space
  • mixup() -- Perform mixup in the semantic relation space

• QLearning.py

  • GNN_env() -- Calculate rewards, perform actions, and update states
  • isTerminal() -- Determine whether the termination conditions have been met

Running the code

1. Install the required dependency packages

2. To get the results on a specific dataset, first run with proper hyperparameters to perform pre-training

python train.py --dataset data_name --setting pre-train

where the data_name is one of the 3 datasets (CCora, BlagCatalog, and Wiki-CS). The pre-trained model will be saved to the corresponding checkpoint folder in ./checkpoint for evaluation.

3. To fine-tune the pre-trained model, run

python train.py --dataset data_name --setting fine-tune --load model_path

where the model_path is the path where the pre-trained model is saved.

4. We provide five compared baselines in this code. They can be configured via the '--setting' arguments:

• Origin: Vanilla backbone models with '--setting raw'
• Over-Sampling: Repeat nodes in the minority classes with '--setting over-sampling'
• Re-weight: Give samples from minority classes a larger weight when calculating the loss with '--setting re-weight'
• SMOTE: Interpolation in the input space with '--setting smote'
• Embed-SMOTE: Perform SMOTE in the intermediate embedding space with '--setting embed_smote'

Use Embed-SMOTE as an example:

python train.py --dataset cora --setting embed_smote

Citation

If you find this project useful for your research, please use the following BibTeX entry.

@inproceedings{wu2023graphmixup,
  title={Graphmixup: Improving class-imbalanced node classification by reinforcement mixup and self-supervised context prediction},
  author={Wu, Lirong and Xia, Jun and Gao, Zhangyang and Lin, Haitao and Tan, Cheng and Li, Stan Z},
  booktitle={Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2022, Grenoble, France, September 19--23, 2022, Proceedings, Part IV},
  pages={519--535},
  year={2023},
  organization={Springer}
}

graphmixup's Issues

关于GAT的结果

您好，我在做对比实验的时候发现编码器为GAT的结果很差，请问是什么原因？