ScaleNet

After paper [Scale Invariance of Graph Neural Networks].

Requirements

This repository has been tested with the following packages:

• Python == 3.9 - 3.12
• PyTorch == 2.1.2
• PyTorch Geometric == 2.4.0
• torch-scatter==2.1.2
• torch-sparse==0.6.18

Performance may vary slightly with different version of Python, GPU, cuda, PyTorch.

Installation Instructions

1. Follow the official instructions to install PyTorch.
2. Install PyTorch Geometric.
3. For pytorch-scatter and pytorch-sparse, download the packages from this link according to your PyTorch, Python, and OS version. Then, use pip to install them.

By following these steps, you can resolve compatibility issues and avoid segmentation faults.

Dataset

Specify the name of the dataset you want to use. The available datasets are categorized as follows:

• Directed Datasets:

  • Assortative Graph:

    • citeseer/
    • cora_ml/
    • WikiCS/
    • telegram/
    • dgl/pubmed

  • Disassortative Graph:

    • WikipediaNetwork/squirrel
    • WikipediaNetwork/chameleon

GNN Models

• GNN baselines:

  • MLP
  • GCN
  • GAT
  • SAGE for GraphSAGE
  • Cheb
  • APPNP

• Hermitian Matrix GNNs:

  • Mag for MagNet
  • Sig for SigMaNet
  • Qua for QuaNet

• Symmetric models:

  • Sym for DGCN

    • 1ym

  • DiG, DiGi2 for DiGCN(ib)

    • 1iG
    • RiG

    In DiG*, 1iG*, RiG*, * can be nothing or exampled as follows**:

    • i2 interception of 2-order edges
    • u3 union of 3-order edges Number 2, 3 can be replaced with any number k>1.

• BiDirectional models:

  • Dir-GNN

How to Run

• (1) To run and get the best performance for each model:

  • On original datasets:

    python3 main.py  --net='ScaleNet' --use_best_hyperparams=1  --Dataset='cora_ml/'
    

    python3 main.py --net='Dir-GNN' --use_best_hyperparams=1   --Dataset='citeseer_npz/'
    
    

  • For imbalanced datasets:

    python3 main.py  --net='ScaleNet' --use_best_hyperparams=1  --Dataset='cora_ml/'   --MakeImbalance   --imb_ratio=100
    

• (2)Run in batches: To run with your own configurations, revise net_nest.h by kicking in all the nets in net_values, all the layers in layer_values, all the datasets in Direct_dataset. Then in terminal, run:

  ./net_nest.h
  

• (3) To compare ScaleNet with the enumeration of the parameters alpha, beta, and gamma, use the following command:

  ./scale.h &
  

• (4) To get performance of removing shared edges with lower-scale graphs:

  • To get performance of AAt-A-At, AtA-A-At, AAt+AtA-A-At ('-' means removing the shared edges with A or At):

    args.differ_AAt=1 args.differ_AA=0

  • To get performance of AA-A-At, AtAt-A-At, AA+AtAt-A-At ('-' means removing the shared edges with A or At):

    args.differ_AA=1

• (5) Wilcoxon test To run the Wilcoxon test on each dataset, execute the corresponding script. For example:

  python3 ./wilcoxon/wilcoxon_cham.py  &
  

License

MIT License

Acknowledgements

The code is implemented based on DiGCN, DirGNNand MagNet.