This repository contains the code for the paper On Calibration of Graph Neural Networks for Node Classification.
The dependencies required to run the code are specified in pyproject.toml. Run poetry install to install the dependencies from poetry.lock, which lists the exact versions used. For more information about Poetry, a tool for dependency management and packaging in Python, see https://python-poetry.org/docs/.
The commands for running the experiments from the paper can be found in run.txt in the corresponding folders.
Defined in args.py.
--dataset: str. Dataset name.
--model: str. Model name.
--num_runs: int. Number of independent runs for each model.
--epochs: int. Number of training epochs.
--early_stopping: bool. Whether early stopping is applied. Note that if no early stopping is used, do not pass this argument to the command line (also not --early_stopping False) since almost anything will be interpreted as True by Python. The same holds for the other boolean arguments.
--patience: int. Number of epochs before doing early stopping.
--tune_wd: bool. Whether a hyperparameter search is conducted for the weight decay value.
--max_search: int. Number of searches for weight decay tuning.
--add_cal_loss: bool. Whether the calibration loss term is added.
--alpha: float. Hyperparameter for the calibration loss.
--lmbda: float. Hyperparameter for the calibration loss.
--num_bins_rbs: int. Number of bins used for the calibration method RBS.