Toward Enhanced Robustness in Unsupervised Graph Representation Learning: A Graph Information Bottleneck Perspective

About

This project is the implementation of the paper "Toward Enhanced Robustness in Unsupervised Graph Representation Learning: A Graph Information Bottleneck Perspectiv".

Dependencies

The following packages are mainly required (along with their dependencies):

• torch==1.12.1
• torch-geometric==2.3.1

Some Python module dependencies are listed in requirements.txt, which can be easily installed with pip:

pip install -r requirements.txt

Usage: Model Training

You can train the robust representations by the main script main.py

Main Script

The help information of the main script main.py is listed as follows:

python main.py -h

usage: main.py [-h] [--dataset DATASET] [--emb_dim EMB_DIM]
           [--nlayers NLAYERS] [--training] [--adv] [--rate RATE]
           [--eps_x EPS_X] [--alpha ALPHA] [--beta BETA] [--device DEVICE]

Process some integers.

optional arguments:
  -h, --help         show this help message and exit
  --dataset DATASET  The required dataset
  --emb_dim EMB_DIM  The embedding dim of the encoder
  --nlayers NLAYERS  The number of layers of the GNN
  --training         Whether train the encoder or load exising encoder
  --adv              Whether evaluate the representations on the adversarial
                     graph
  --rate RATE        The edge attack ratio
  --eps_x EPS_X      The features attack ratio
  --alpha ALPHA      The hyper-parameter to control the weight of the
                     adversarial term
  --beta BETA        The hyper-parameter to control the weight of the KL
                     divergence term
  --device DEVICE    The device you use

Demo

Then a demo script is available by calling main.py, as the following:

python main.py --dataset cora --alpha 0.9 --beta 0.1 --rate 0.2 --eps_x 0.001 --training --device 1

Usage: Evaluation

To evaluate the robustness on the adversarial graph, you need to generate the adversarial graphs first. We provide the script generate_adv.py to generate the adversarial graphs by Deeprobust.

The help information of the main script generate_adv.py is listed as follows:

python generate_adv.py -h

usage: generate_adv.py [-h] [--dataset DATASET] [--rate RATE] [--eps_x EPS_X]
                       [--device DEVICE]

optional arguments:
  -h, --help         show this help message and exit
  --dataset DATASET  The dataset required.
  --rate RATE        The ratio for edge perturbations.
  --eps_x EPS_X      The ratio for feature perturbations.
  --device DEVICE    The devive you use.

Demo

Then a demo script is available by calling generate_adv.py, as the following:

python generate_adv.py --dataset cora

Then, you can evaluate the robustness of RGIB with the adversarial graphs:

python main.py --dataset cora --alpha 0.9 --beta 0.1 --rate 0.2 --eps_x 0.001 --adv --device 1