This repository contains the code for GISST (Graph Neural Networks Including Sparse Interpretability). This method provides important node feature and sub-graph edge interpretation in end-to-end training. Please read example_gisst.ipynb for a quick tutorial on how to use the APIs in this repository.
GISST is developed by Chris Lin, Marylens Hernandez, Gerald Sun, Krishna Bulusu, and Jonathan Dry.
The name SIG was used during the development of this repository. The method was later renamed GISST (Graph Neural Networks Including Sparse Interpretability ). The preprint of GISST is available on arXiv.
The developers are updating GISST for a stable version of PyTorch Geometric.
If using this repository for the first time, execute setup.sh to set up additional project sub-directories and preprocess real-world datasets (Mutagenicity, REDDIT-BINARY, and PROTEINS).
Synthetic datasets are included in graphs. After preprocessing through setup.sh, real-world datasets will also be added to your local graphs directory.
Models that have been tuned and trained are included in output. The directory of each trained model contains the following files.
configs.yaml: configuration file with the hyperparameters that were searched through.best_results.txt: classification results of the trained model.best_hyperparams.txt: the optimal set of hyperparameters selected for the trained model.best_state_dict.pt: the parameters/weights of the trained model.
With best_hyperparams.txt and best_state_dict.pt, the trained model can be restored in Pytorch (see explain_node.py for example).