GIB
This repository reproduces the results in the paper Graph Information Bottleneck (Tailin Wu *, Hongyu Ren *, Pan Li, Jure Leskovec, NeurIPS 2020), whose objective is to learn minimal sufficient structural and feature information using GNNs, which improves the robustness of GNNs.
Representation learning on graphs with graph neural networks (GNNs) is a challenging task. Previous work has shown that GNNs are susceptible to adversarial attack. We here introduce Graph Information Bottleneck (GIB), which learns representation that is maximally informative about the target to predict while using minimal sufficient information of the input data. Concretely, the GIB principle regularizes the representation of the node features as well as the graph structure so that it increases the robustness of GNNs. For more information, see our paper Graph Information Bottleneck (Wu et al. 2020), and our project website at http://snap.stanford.edu/gib/.
Installation
First clone the directory. Then run the following command to initialize the submodules:
git submodule init; git submodule update(If showing error of no permission, need to first add a new SSH key to your GitHub account.)
The repository also has the following dependencies, and please refer to the respective page to install:
- PyTorch >= 1.4.0
- pytorch_geometric >=1.4.0,<=1.5.1
Additional requirements are in requirements.txt, which can be installed via pip install -r requirements.txt.
After installing the dependencies, cd to the directory "DeepRobust/", and install it by running:
pip install -e .Usage
The main experiment files are:
- Node-level experiments: experiments/GIB_node_exp.ipynb
- Node-level attacks: experiments/GIB_node_attack_exp.ipynb
which can be run via command line or in Jupyter notebook.
The result files are saved under the "results/" folder.
The definition of GIB-GAT, GAT, GCN are in experiments/GIB_node_model.ipynb.
The analysis script is experiments/GIB_node_analysis.ipynb.
Run adversarial attack experiments
To run multiple attack experiments each with a different hyperparameter combination, run "run_exp/run_nettack_grid.py" by e.g.
python run_exp/run_nettack_grid.py ${Assign_ID} ${GPU_ID}where each integer ${Assign_ID} (0 to M-1) maps to a hyperparameter setting (M is the total number of hyperparameter settings), and ${GPU_ID} is the ID (e.g. 0, 1, 2) of CUDA driver (set to False if using CPU).
Alternatively, to run a single attack experiment, use "run_exp/run_nettack.py". Below are the commands that produce the adversarial attack results in the paper (For node feature attacks, see the README in run_exp/). For the args, the "exp_id" and "date_time" are used to name the folder "{}_{}".format(exp_id, date_time) in which the results will be saved in. "gpuid" can also be set in a custom way. For each experiment, need to go over seeds of 0, 1, 2, 3, 4 then perform analysis, where in the following for brevity we only provide --seed=0. Also note that the following "data_type" all have suffix of "-bool", which makes the feature Boolean as required by Netteck. After running each experiment, use the script experiments/GIB_node_analysis.ipynb (Section 2) to perform analysis and obtain results.
Cora with GIB-Cat:
python run_exp/run_nettack.py --exp_id=Cora-GIB-Cat --data_type=Cora-bool --model_type=GAT --beta1=0.001 --beta2=0.01 --struct_dropout_mode='\("DNsampling","multi-categorical-sum",1,3,2\)' --seed=0 --gpuid=0Cora with GIB-Bern:
python run_exp/run_nettack.py --exp_id=Cora-GIB-Bern --data_type=Cora-bool --model_type=GAT --beta1=0.001 --beta2=0.01 --struct_dropout_mode='\("DNsampling","Bernoulli",0.1,0.5,"norm",2\)' --seed=0 --gpuid=0Pubmed with GIB-Cat:
python run_exp/run_nettack.py --exp_id=Pubmed-GIB-Cat --data_type=Pubmed-bool --model_type=GAT --beta1=0.001 --beta2=0.01 --struct_dropout_mode='\("DNsampling","multi-categorical-sum",1,3,2\)' --seed=0 --gpuid=0Pubmed with GIB-Bern:
python run_exp/run_nettack.py --exp_id=Pubmed-GIB-Bern --data_type=Pubmed-bool --model_type=GAT --beta1=0.001 --beta2=0.01 --struct_dropout_mode='\("DNsampling","Bernoulli",0.1,0.5,"norm",2\)' --seed=0 --gpuid=0Citeseer with GIB-Cat:
python run_exp/run_nettack.py --exp_id=Citeseer-GIB-Cat --data_type=citeseer-bool --model_type=GAT --beta1=0.001 --beta2=0.01 --struct_dropout_mode='\("DNsampling","multi-categorical-sum",0.1,2,2\)' --seed=0 --gpuid=0Citeseer with GIB-Bern:
python run_exp/run_nettack.py --exp_id=Citeseer-GIB-Bern --data_type=citeseer-bool --model_type=GAT --beta1=0.001 --beta2=0.01 --struct_dropout_mode='\("DNsampling","Bernoulli",0.05,0.5,"norm",2\)' --seed=0 --gpuid=0Other baselines:
Cora with GAT:
python run_exp/run_nettack.py --exp_id=Cora-GAT --data_type=Cora-bool --model_type=GAT --beta1=-1 --beta2=-1 --struct_dropout_mode='\("standard",0.6\)' --seed=0 --gpuid=0Cora with GCN:
python run_exp/run_nettack.py --exp_id=Cora-GCN --data_type=Cora-bool --model_type=GCN --beta1=-1 --beta2=-1 --seed=0 --gpuid=0Cora with GCNJaccard:
python run_exp/run_nettack.py --exp_id=Cora-GCNJaccard --data_type=Cora-bool --model_type=GCNJaccard --beta1=-1 --beta2=-1 --latent_size=16 --lr=1e-2 --weight_decay=5e-4 --threshold=0.05 --seed=0 --gpuid=0Cora with RGCN:
python run_exp/run_nettack.py --exp_id=Cora-RGCN --data_type=Cora-bool --model_type=RGCN --beta1=5e-4 --beta2=-1 --latent_size=64 --lr=1e-2 --weight_decay=5e-4 --gamma=0.3 --seed=0 --gpuid=0Pubmed with GAT:
python run_exp/run_nettack.py --exp_id=Pubmed-GAT --data_type=Pubmed-bool --model_type=GAT --beta1=-1 --beta2=-1 --struct_dropout_mode='\("standard",0.6\)' --seed=0 --gpuid=0Pubmed with GCN:
python run_exp/run_nettack.py --exp_id=Pubmed-GCN --data_type=Pubmed-bool --model_type=GCN --beta1=-1 --beta2=-1 --seed=0 --gpuid=0Pubmed with GCNJaccard:
python run_exp/run_nettack.py --exp_id=Pubmed-GCNJaccard --data_type=Pubmed-bool --model_type=GCNJaccard --beta1=-1 --beta2=-1 --latent_size=16 --lr=1e-2 --weight_decay=5e-4 --threshold=0.05 --seed=0 --gpuid=0Pubmed with RGCN:
python run_exp/run_nettack.py --exp_id=Pubmed-RGCN --data_type=Pubmed-bool --model_type=RGCN --beta1=5e-4 --beta2=-1 --latent_size=16 --lr=1e-2 --weight_decay=5e-4 --gamma=0.1 --seed=0 --gpuid=0Citeseer with GAT:
python run_exp/run_nettack.py --exp_id=Citeseer-GAT --data_type=citeseer-bool --model_type=GAT --beta1=-1 --beta2=-1 --struct_dropout_mode='\("standard",0.6\)' --seed=0 --gpuid=0Citeseer with GCN:
python run_exp/run_nettack.py --exp_id=Citeseer-GCN --data_type=citeseer-bool --model_type=GCN --beta1=-1 --beta2=-1 --seed=0 --gpuid=0Citeseer with GCNJaccard:
python run_exp/run_nettack.py --exp_id=Citeseer-GCNJaccard --data_type=citeseer-bool --model_type=GCNJaccard --beta1=-1 --beta2=-1 --latent_size=16 --lr=1e-2 --weight_decay=5e-4 --threshold=0.05 --seed=0 --gpuid=0Citeseer with RGCN:
python run_exp/run_nettack.py --exp_id=Citeseer-RGCN --data_type=citeseer-bool --model_type=RGCN --beta1=5e-4 --beta2=-1 --latent_size=64 --lr=1e-2 --weight_decay=5e-4 --gamma=0.3 --seed=0 --gpuid=0Ablation study:
Cora with XIB:
python run_exp/run_nettack.py --exp_id=Cora-XIB --data_type=Cora-bool --model_type=GAT --beta1=0.001 --beta2=-1 --struct_dropout_mode='\("standard",0.6,2\)' --seed=0 --gpuid=0Cora with AIB-Cat:
python run_exp/run_nettack.py --exp_id=Cora-AIB-Cat --data_type=Cora-bool --model_type=GAT --beta1=-1 --beta2=0.01 --struct_dropout_mode='\("DNsampling","multi-categorical-sum",1,3,2\)' --seed=0 --gpuid=0Cora with AIB-Bern:
python run_exp/run_nettack.py --exp_id=Cora-AIB-Bern --data_type=Cora-bool --model_type=GAT --beta1=-1 --beta2=0.01 --struct_dropout_mode='\("DNsampling","Bernoulli",0.1,0.5,"norm",2\)' --seed=0 --gpuid=0Citation
If you compare with, build on, or use aspects of the Graph Information Bottleneck, please cite the following:
@inproceedings{wu2020graph,
title={Graph Information Bottleneck},
author={Wu, Tailin and Ren, Hongyu and Li, Pan and Leskovec, Jure},
booktitle={Neural Information Processing Systems},
year={2020},
}