Developing Graph Neural Networks (GNNs) for heterogeneous graphs and overseeing the explainability.
pip install -r requirements.txt
It is recommended to run on Datahub instance since some of the packages might not be recognized while running on the local machine.
gnn/captum_explainer.py contains the implementation of GNN model and the explainer for all four attribution methods. The remaining folders are helpers.
This script generates visualizations for the feature importance using each of the methods which have been stored in the results folder.
- PyTorch Geomtric does not yet support Captum evaluation metrics to evaluate the truthfulness of the explanations for regression tasks.
- Relbench is a novel framework, support is limited and it is still in development phase.