RS-GNN

An offical PyTorch implementation of "Towards Robust Graph Neural Networks for Noisy Graphs with Sparse Labels" (WSDM 2022). [paper]

Overview

• ./models: This directory contains the model of RSGNN.
• ./dataset.py: This is the code to load datasets and perturbed adjacency matrix.
• ./data: The pre-perturbed adjacency matrices of the datasets are stored here.
• ./scripts: It contains the scripts to reproduce the major reuslts of our paper.
• ./generate_attack.py: An example code of obtaining the perturbed dataset. To run this code, it is required to install DeepRobust
• ./train_RSGNN.py: The program to train RSGNN model.

Dataset

The original Cora, Cora-ML, Citeseer, and Pubmed will be automatically downloaded to ./data. The val and test indices are the same as nettack settings.

For the perturbed adjacency matrix, it is stored as: ./data/{label_rate}/{dataset}_{attack_method}_adj_{ptb_rate}.npz.

Requirements

torch==1.7.1
torch-geometric==1.7.2 

Experiments

To reproduce the performance in the paper, you can run the bash files in the .\scripts. For example, to get results on cora datasets

bash scripts\train_cora.sh

Cite

If you find this repo to be useful, please cite our paper. Thank you.

@article{dai2022towards,
  title={Towards Robust Graph Neural Networks for Noisy Graphs with Sparse Labels},
  author={Dai, Enyan and Jin, Wei and Liu, Hui and Wang, Suhang},
  journal={WSDM},
  year={2022}
}