This code uses:

- python 3.8
- torch 1.11.0
- torch-geometric 2.0.4
- PyGCL 0.1.2

For install a complete list of dependecies, use pip install -r requirements.txt.

mkdir data
mkdir logs
cd datagen
./create_dataset.sh #create a synthetic dataset to evaluate on.
cd ../examples 
./run_method.sh #train a graph SSL method the generated dataset.
./run_svc.ch #evaluate over different style ratios.
./run_compute_invariance_sep.sh #compute the separability vs. invariance scores.

We provide an example script for generating a synthetic dataset in datagen/create_dataset.sh. With default settings, dataset files will be saved in, ./data/. Augmentations designed for this synthetic dataset are provided in:

• edge_removing_caa.py: Content-aware, oracle edge dropping augmentation. This augmentation will only drop edges that do not belong to "content" or inserted "motif" in each sample.

• edge_removing_gga.py: Generic edge dropping augmentation. This augmentation randomly drops a percentage of the total edges in each sample.

We provide reference implementations for Augmentation Augmented Graph Autoencoder, Graph Autoencoder, GraphCL, SimGRACE and SpecCL. examples/run_method.sh provides an example script for running SpecCL. (Running other methods is similar.)

• examples/compute_svc.py: given a trained checkpoint, compute the accuracy over different style vs. content ratios. See run_svc.sh for an example.
• examples/compute_invariance_sep.py: given a trained checkpoint, compute invariance and separability scores. See run_compute_invariance.sh for an example.

This code is inspired by and makes us of several great code bases. We especially thank the authors of PyGCL, GraphCL, AD-GCL, and GNNExplainer.