Label Noise-Resistant Graph Neural Network

Offical pytorch implementation of proposed framework NRGNN and Compared Methods in "NRGNN: Learning a Label Noise-Resistant Graph Neural Network on Sparsely and Noisily Labeled Graphs" (KDD 2021). If you find this repo to be useful, please cite our paper. Thank you.

@article{dai2021nrgnn,
  title={NRGNN: Learning a Label Noise-Resistant Graph Neural Network on Sparsely and Noisily Labeled Graphs},
  author={Dai, Enyan and Aggarwal, Charu and Wang, Suhang},
  journal={SIGKDD},
  year={2021}
}

Content

• Label Noise-Resistant Graph Neural Network
  • Content
  • 1. Requirements
  • 2. NRGNN
    • 2.1 Introduction
    • Reproduc the Results
  • 3. Compared Methods (to test)
    • Co-teaching+
    • D-GNN
    • LafAK (CP)
  • 4. Dataset

1. Requirements

torch==1.7.1
torch-geometric==1.7.1

The packages can be installed by directly run the commands in install.sh by

bash install.sh

2. NRGNN

2.1 Introduction

we propose to link the unlabeled nodes with labeled nodes of high feature similarity to bring more clean label information. Furthermore, accurate pseudo labels could be obtained by this strategy to provide more supervision and further reduce the effects of label noise.

Reproduc the Results

An example of training NRGNN:

python train_NRGNN.py \
    --dataset cora \
    --seed 11 \
    --t_small 0.1 \
    --alpha 0.03\
    --lr 0.001 \
    --epochs 200 \
    --n_p -1 \
    --p_u 0.8 \
    --label_rate 0.05 \
    --ptb_rate 0.2 \
    --noise uniform

All the hyper-parameters settings for the datasets are included in train_NRGNN.sh. To reproduce the performance reported in the paper, you can run the bash file:

bash train_NRGNN.sh

3. Compared Methods (to test)

Co-teaching+

From Yu, Xingrui, et al. "How does disagreement help generalization against label corruption?." [model, trianing_example]

D-GNN

D-GNN is based on S-model from NT, Hoang, Choong Jun Jin, and Tsuyoshi Murata. "Learning graph neural networks with noisy labels." [model, trianing_example]

LafAK (CP)

From Zhang, Mengmei, et al. "Adversarial label-flipping attack and defense for graph neural networks." [model, trianing_example] (Note that it is required to run deepwalk to obtain node embedding for the code.)

Our reimplemention of these methods are in ./models. Their example training codes are in ./baseline.

4. Dataset

The DBLP can be automatically downloaded to ./data through torch-geometric API. Other datasets are listed in ./data