Extension of INVASE to GNNs (work in progress).
This directory contains implementations of INVASE framework for the following applications on graph input data.
- Instance-wise feature and node selection
- Prediction with instance-wise feature and node selection
To run the pipeline for training and evaluation on INVASE-GNN framwork on the MUTAG dataset,
simply run python main.py. Experiment and model hyperparameters can be adjusted using command line flags. All code is in PyTorch, with GNNs implemented with PyTorch Geometric
In addition, GNNExplainer.py contains an example of the post-hoc graph explanation approach introduced in GNNExplainer.
$ python3 main.py --model-type INVASE --task mutag \
--node-lamda 0.01 --fea-lamda 0.01 --l2 0.0 --dropout 0.0 \
--actor-h-dim 64 --critic-h-dim 64 --n-layer 3 \
--batch-size 256 --epochs 300 --lr 0.01 \
--run-id 1 - Comparison to other methods
- Synthetic data experiments
- Node classification feature selection
- Compare edge vs node masking
Original paper:
Jinsung Yoon, James Jordon, Mihaela van der Schaar,
"INVASE: Instance-wise Variable Selection using Neural Networks,"
International Conference on Learning Representations (ICLR), 2019.
(https://openreview.net/forum?id=BJg_roAcK7)
This code was built on top of the INVASE code repository, and some inspiration was taken from invase-pytorch.