Invertible Graph Neural Network iGNN

Implementation and experiments based on the paper Invertible Neural Network for Graph Prediction, accepted at the IEEE Journal on Selected Areas in Information Theory---Deep Learning for Inverse Problems.

Citation:

@ARTICLE{9950057,
  author={Xu, Chen and Cheng, Xiuyuan and Xie, Yao},
  journal={IEEE Journal on Selected Areas in Information Theory}, 
  title={Invertible Neural Networks for Graph Prediction}, 
  year={2022},
  volume={3},
  number={3},
  pages={454-467},
  doi={10.1109/JSAIT.2022.3221864}}

• Please see example.ipynb regarding how to use the method.
• The movie below visualizes how iGNN transports original densities $X|Y$ of the three-moon dataset to their corresponding $H|Y$. The top row plots the Wasserstein-2 penalty at each block, where larger values indicate more drastic amount of movement by the block.