https://github.com/CheriseZhu/GSSNN

The implementation of our AAAI 2020 paper "GSSNN: Graph Smoothing Splines Neural Network".
https://github.com/CheriseZhu/GSSNN

Last synced: 19 days ago
JSON representation

The implementation of our AAAI 2020 paper "GSSNN: Graph Smoothing Splines Neural Network".

• Host: GitHub
• URL: https://github.com/CheriseZhu/GSSNN
• Owner: CheriseZhu
• Created: 2019-11-14T12:30:17.000Z (about 5 years ago)
• Default Branch: master
• Last Pushed: 2020-07-30T12:32:45.000Z (over 4 years ago)
• Last Synced: 2024-08-01T17:24:46.628Z (4 months ago)
• Language: Python
• Size: 1.73 MB
• Stars: 20
• Watchers: 3
• Forks: 7
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-graph-classification - [Python Reference

README

        
# GSSNN

The implementation of our AAAI 2020 paper "[GSSNN: Graph Smoothing Splines Neural Network](https://www.researchgate.net/publication/337548602_GSSNN_Graph_Smoothing_Splines_Neural_Networks)". 

# Requirements

python == 3.6.2


torch == 1.1.0


numpy == 1.16.4


scipy == 1.2.0


networkx == 2.2


torch_scatter == 1.3.0


torch_geometric == 1.3.0

# How to use

  ### Dataset

  The data folder includes our propocessed data for training and testing. 


  The orginal datasets can be founded from [here](https://ls11-www.cs.tu-dortmund.de/people/morris/graphkerneldatasets).

  ### Model

  The model folder includes our proposed model "GSSNN".


  The model/utils folder includes graph utils and Scaled Smoothing Splines module used in model.


  The model/process_data file processes data and computes the graph centrality.


  The torch_geometeric/nn/pool folder includes the designed NodeImportance layer used in model.


  The torch_geometeric/nn/conv folder includes the convolutional layers used in model provides by [torch_geometeric](https://github.com/rusty1s/pytorch_geometric) library.


  

  ### Prameters Setting

  dim: the hidden dimension of node feature


  conv_layer: the number of convolutional layer


  ss_layer: the number of smoothing splines layer


  Mi: the number of knot used in smoothing splines layer i


  epsilon: used in smoothing splines to guarantee the denominator non-zero


  add_knot: whether to consider the important nodes features as residual connection to the graph-level representation


  

  ### Training/Testing

  ```

  cd model

  python process_data.py --dataset MUTAG

  python GSSNN.py --dataset MUTAG --batch_size 128 --lr 0.01 --weight_decay 5e-4 --dim 32 --conv_layer 3 --ss_layer 2 --M1 5 --M2 5 --epsilon 1e-6 --add_knot True --epoch 100

  python GSSNN_10_folds.py --dataset MUTAG --batch_size 128 --lr 0.01 --weight_decay 5e-4 --dim 32 --conv_layer 3 --ss_layer 2 --M1 5 --M2 5 --epsilon 1e-6 --add_knot True --epoch 100

  ```

  

  

# Citation

```

@inproceedings{zhu2020GSSNN

author={Shichao Zhu and Lewei Zhou and Shirui Pan and Chuan Zhou and Guiying Yan and Bin Wang},

title={GSSNN: Graph Smoothing Splines Neural Network},

journal={Proceedings of the 34th AAAI Conference on Artificial Intelligence},

year={2020}

}

```