TianBian95 / pi-gnn

Noise-robust Graph Learning by Estimating and Leveraging Pairwise Interactions, TMLR, 06/2023.

1 Package Dependency

We provide a list of reference versions for all packages in the requirements.txt, pip can be used to install these packages:

pip3 install -r requirements.txt

2 Run the code

We provide some scripts in the ./scripts folder to run our code. For example, you can run the PI-GNN model by the command bash ./scripts/PI-GNN.sh. Below, we will explain the meanings of different commands in these scripts in detail.

2.1 Variables

First, we need to define some common variables:

dataset=cora      #Datasets for evaluation: cora, cite, pub, wiki, ogbn-arxiv
type=gcn          #Backbone: gcn, gat, sage
cor_type=uniform  #The type of label noise: uniform, flip
cor_prob=0.2      #The ratio of label noise: [0.0, 1]
epochs=400        #Training epochs

More variables will be defined in specific scripts.

2.2 Classical graph learning models

For classical graph learning models (e.g., GCN, GAT, SAGE), run:

python3 ./model/GCN.py --dataset $dataset --type $type --corruption_type $cor_type --corruption_prob $cor_prob --epochs $epochs

For large-scale dataset, such as ogbn-arxiv, run:

python3 ./model_bigdata/GCN.py --dataset $dataset --type $type --corruption_type $cor_type --corruption_prob $cor_prob --epochs $epochs

2.3 PI-GNN w/o pc (predictive confidence)

For PI-GNN without predictive confidence, i.e., PI-GNN trained with the node connectivity, run:

python3 ./model/PI-GCNwopc.py --dataset $dataset --type $type --corruption_type $cor_type --corruption_prob $cor_prob --epochs $epochs

For large-scale dataset, such as ogbn-arxiv, run:

python3 ./model_bigdata/PI-GCNwopc.py --dataset $dataset --type $type --corruption_type $cor_type --corruption_prob $cor_prob --epochs $epochs

2.4 PI-GNN

For PI-GNN, run:

python3 ./model/PI-GCN.py --dataset $dataset --type $type --corruption_type $cor_type --corruption_prob $cor_prob --epochs $epochs

For large-scale dataset, such as ogbn-arxiv, run:

python3 ./model_bigdata/PI-GCN.py --dataset $dataset --type $type --corruption_type $cor_type --corruption_prob $cor_prob --epochs $epochs

3 Dataset

When you execute the above code, the PyTorch Geometric package will automatically download the corresponding raw data in the ./data/$dataset/raw/ folder, and save the processed data in the ./data/$dataset/processed/ folder. We also provide the raw data in data/ folder in case of data version inconsistency or failure to download due to network problems.

Citation

If you found any part of this code is useful in your research, please consider citing our paper:

@article{du2023noise,
    title={Noise-robust Graph Learning by Estimating and Leveraging Pairwise Interactions},
    author={Du, Xuefeng and Bian, Tian and Rong, Yu and Han, Bo and Liu, Tongliang and Xu, Tingyang and Huang, Wenbing and Li, Yixuan and Huang, Junzhou},
    journal={Transactions on Machine Learning Research},
    year={2023}
}