This repo contains the implementation of the model proposed
in SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks.
ogbn-arxiv dataset will be loaded automatically, while Cora, Citeseer, and Pubmed are included in the GCN
package, available here. Place the relevant files in the folder
data_tf.
Pythonversion 3.7.2Numpyversion 1.18.5PyTorchversion 1.5.1DGLversion 0.5.2sklearnversion 0.21.3scipyversion 1.2.1torch-geometric1.6.1ogbversion 1.2.3
To train the models, you need a machine with a GPU.
To install the dependencies, it is recommended to use a virtual environment. You can create a virtual environment and install all the dependencies with the following command:
conda env create -f environment.ymlThe file requirements.txt was written for CUDA 9.2 and Linux so you may need to adapt it to your infrastructure.
To run the model you should define the following parameters:
dataset: The dataset you want to run the model onntrials: number of runsepochs_adj: number of epochsepochs: number of epochs for GNN_C (used for knn_gcn and 2step learning of the model)lr_adj: learning rate of GNN_DAElr: learning rate of GNN_Cw_decay_adj: l2 regularization parameter for GNN_DAEw_decay: l2 regularization parameter for GNN_Cnlayers_adj: number of layers for GNN_DAEnlayers: number of layers for GNN_Chidden_adj: hidden size of GNN_DAEhidden: hidden size of GNN_Cdropout1: dropout rate for GNN_DAEdropout2: dropout rate for GNN_Cdropout_adj1: dropout rate on adjacency matrix for GNN_DAEdropout_adj2: dropout rate on adjacency matrix for GNN_Cdropout2: dropout rate for GNN_Ck: k for knn initialization with knnlambda_: weight of loss of GNN_DAEnr: ratio of zeros to ones to mask out for binary featuresratio: ratio of ones to mask out for binary features and ratio of features to mask out for real values featuresmodel: model to run (choices are end2end, knn_gcn, or 2step)sparse: whether to make the adjacency sparse and run operations on sparse modegen_mode: identifies the graph generatornon_linearity: non-linearity to apply on the adjacency matrixmlp_act: activation function to use for the mlp graph generatormlp_h: hidden size of the mlp graph generatornoise: type of noise to add to features (mask or normal)loss: type of GNN_DAE loss (mse or bce)epoch_d: epochs_adj / epoch2 of the epochs will be used for training GNN_DAEhalf_val_as_train: use half of validation for train to get Cora390 and Citeseer370
In order to reproduce the results presented in the paper, you should run the following commands:
Run the following command:
python main.py -dataset cora -ntrials 10 -epochs_adj 2000 -lr 0.001 -lr_adj 0.01 -w_decay 0.0005 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 512 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.5 -dropout_adj2 0.25 -k 30 -lambda_ 10.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 0 -non_linearity elu -epoch_d 5Run the following command:
python main.py -dataset cora -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.001 -w_decay 0.0005 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 512 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.5 -k 20 -lambda_ 10.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 1 -non_linearity relu -mlp_h 1433 -mlp_act relu -epoch_d 5Run the following command:
python main.py -dataset cora -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.001 -w_decay 0.05 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 512 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.5 -k 15 -lambda_ 10.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 2 -non_linearity relu -mlp_act relu -epoch_d 5Run the following command:
python main.py -dataset citeseer -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.01 -w_decay 0.05 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 1024 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.4 -dropout_adj2 0.4 -k 30 -lambda_ 1.0 -nr 1 -ratio 10 -model end2end -sparse 0 -gen_mode 0 -non_linearity elu -epoch_d 5Run the following command:
python main.py -dataset citeseer -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.001 -w_decay 0.0005 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 1024 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.5 -k 30 -lambda_ 10.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 1 -non_linearity relu -mlp_act relu -mlp_h 3703 -epoch_d 5Run the following command:
python main.py -dataset citeseer -ntrials 10 -epochs_adj 2000 -lr 0.001 -lr_adj 0.01 -w_decay 0.05 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 1024 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.5 -dropout_adj2 0.5 -k 20 -lambda_ 10.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 2 -non_linearity relu -mlp_act tanh -epoch_d 5Run the following command:
python main.py -dataset cora -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.01 -w_decay 0.0005 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 512 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.5 -k 20 -lambda_ 100.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 0 -non_linearity elu -epoch_d 5 -half_val_as_train 1Run the following command:
python main.py -dataset cora -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.001 -w_decay 0.0005 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 512 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.5 -k 20 -lambda_ 10.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 1 -non_linearity relu -mlp_h 1433 -mlp_act relu -epoch_d 5 -half_val_as_train 1Run the following command:
python main.py -dataset cora -ntrials 10 -epochs_adj 2000 -lr 0.001 -lr_adj 0.001 -w_decay 0.0005 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 512 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.5 -k 20 -lambda_ 10.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 2 -non_linearity relu -mlp_act relu -epoch_d 5 -half_val_as_train 1Run the following command:
python main.py -dataset citeseer -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.01 -w_decay 0.05 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 1024 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.5 -dropout_adj2 0.5 -k 30 -lambda_ 1.0 -nr 1 -ratio 10 -model end2end -sparse 0 -gen_mode 0 -non_linearity elu -epoch_d 5 -half_val_as_train 1Run the following command:
python main.py -dataset citeseer -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.001 -w_decay 0.0005 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 1024 -dropout1 0.25 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.5 -k 30 -lambda_ 10.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 1 -non_linearity relu -mlp_act tanh -mlp_h 3703 -epoch_d 5 -half_val_as_train 1Run the following command:
python main.py -dataset citeseer -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.01 -w_decay 0.05 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 1024 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.5 -k 20 -lambda_ 10.0 -nr 5 -ratio 10 -model end2end -sparse 0 -gen_mode 2 -non_linearity relu -mlp_act tanh -epoch_d 5 -half_val_as_train 1Run the following command:
python main.py -dataset pubmed -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.01 -w_decay 0.0005 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 128 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.5 -dropout_adj2 0.5 -k 15 -lambda_ 10.0 -nr 5 -ratio 20 -model end2end -gen_mode 1 -non_linearity relu -mlp_h 500 -mlp_act relu -epoch_d 5 -sparse 1Run the following command:
python main.py -dataset pubmed -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.01 -w_decay 0.0005 -nlayers 2 -nlayers_adj 2 -hidden 32 -hidden_adj 128 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.25 -k 15 -lambda_ 100.0 -nr 5 -ratio 20 -model end2end -sparse 0 -gen_mode 2 -non_linearity relu -mlp_act tanh -epoch_d 5 -sparse 1Run the following command:
python main.py -dataset ogbn-arxiv -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.001 -w_decay 0.0 -nlayers 2 -nlayers_adj 2 -hidden 256 -hidden_adj 256 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.25 -dropout_adj2 0.5 -k 15 -lambda_ 10.0 -nr 5 -ratio 100 -model end2end -sparse 0 -gen_mode 1 -non_linearity relu -mlp_h 128 -mlp_act relu -epoch_d 2001 -sparse 1 -loss mse -noise maskRun the following command:
python main.py -dataset ogbn-arxiv -ntrials 10 -epochs_adj 2000 -lr 0.01 -lr_adj 0.001 -w_decay 0.0 -nlayers 2 -nlayers_adj 2 -hidden 256 -hidden_adj 256 -dropout1 0.5 -dropout2 0.5 -dropout_adj1 0.5 -dropout_adj2 0.25 -k 15 -lambda_ 10.0 -nr 5 -ratio 100 -model end2end -sparse 0 -gen_mode 2 -non_linearity relu -mlp_act relu -epoch_d 2001 -sparse 1 -loss mse -noise normalIf you use this package for published work, please cite the following:
@inproceedigs{fatemi2021slaps,
title={SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks},
author={Fatemi, Bahare and Asri, Layla El and Kazemi, Seyed Mehran},
booktitle={Advances in Neural Information Processing Systems},
year={2021}
}