This codebase has been used to generate all results based on the real-world datasets in the ICLR 2023 paper Revisiting Robustness in Graph Machine Learning.
The repository requires the python package rgnn_at_scale. Please follow the installation instructions to install the package as well as other required packages. The custom coda kernels of rgnn_at_scale are not required.
The experiment experiment_train_inductive.py performs inductive training for the in the config specified models. The corresponding config files can be found in config/train
To train the models run
python script_execute_experiment.py --config-file 'config/evaluate/cora_ml_and_citeseer.yaml'The trained models are stored in the cache folder and the output is logged into ./output.
The experiment experiment_evaluate_overrobustness.py evaluates the overrobustness for the models saved in cache. The corresponding config files can be found in config/evaluate
To run the evaluation run
bash python script_execute_experiment.py --config-file 'config/evaluate/cora_ml_and_citeseer.yaml'
The output is logged into ./output.
This codebase contains code snippets from the following repositories:
We thank the authors for making their code public and the development team of PyTorch Geometric.