Graph Structure Self-Contrasting (GSSC)

This is a PyTorch implementation of Graph Structure Self-Contrasting (GSSC), and the code includes the following modules:

• Dataset Loader (Cora, Citeseer, Actor, Coauthor-CS, and Coauthor-Phy)

• Various baselines (GCN, GraphSage, GAT, SGC, APPNP, and DAGNN compared in this paper)

• Neighborhood sparsification (NeighSparse) and Neighborhood self-contrasting (NeighContrast) network

• Bi-level optimization framework

• Training and evaluation paradigm

Main Requirements

• networkx==2.5
• numpy==1.19.2
• dgl==0.6.1
• torch==1.6.0

Description

• main.py

  • main() -- Train the model for node classification task with NeighContrast network on five datasets.

• model.py

  • MLP_GNN() -- A pure MLP-based architecture ans two prediction heads.
  • EdgeSampler() -- Neighborhood Sparsification (NeighSparse) Network.
  • MixupScale() -- Learning Sampling Coefficients for interpolation between the target node and its ineighborhood nodes.

• dataset.py

  • dataloader() -- Load five datasets as well as their variants with different label rates, label noise, and struture disturbance.

• utils.py

  • SetSeed() -- Set seeds for reproducible results.

Running the code

1. Install the required dependency packages

2. To get the results on a specific dataset, please run with proper hyperparameters:

python main.py --dataset data_name

where the data_name is one of the 5 datasets (Cora, Citeseer, Actor, Coauthor-CS, and Coauthor-Phy) . Use Cora dataset an example:

python main.py --dataset cora

3. To get the results on datasets with different label rates, label noise, and struture disturbance, please run with:

python main.py --label_mode

where the label_mode denotes different dataset variants.

(1) -1: training with 1 labels per class;

(2) -3: training with 3 labels per class;

(3) -5: training with 5 labels per class;

(4) -10: training with 5 labels per class;

(5) -25: training with structure perturbation ratio 5%;

(6) -30: training with structure perturbation ratio 10%;

(7) -40: training with structure perturbation ratio 20%;

(8) -50: training with structure perturbation ratio 30%;

(9) -51: training with symmetric label noise ratio 20%;

(10) -52: training with symmetric label noise ratio 40%;

(11) -53: training with symmetric label noise ratio 60%;

(12) -54: training with asymmetric label noise ratio 20%;

(13) -55: training with asymmetric label noise ratio 40%;

(14) -56: training with asymmetric label noise ratio 60%.

Citation

If you are interested in our repository and our paper, please cite the following paper:

@article{wu2024learning,
  title={Learning to Model Graph Structural Information on MLPs via Graph Structure Self-Contrasting},
  author={Wu, Lirong and Lin, Haitao and Zhao, Guojiang and Tan, Cheng and Li, Stan Z},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2024},
  publisher={IEEE}
}