Graph Highway Networks in JAX

This is a non-official implementation of the recent GHNets in JAX. The code contains the Graph Highway Networks definition with the three types of node feature infusion. More details in the original paper Graph Highway Networks.

Usage

Run

python train.py

to train a model on the Cora dataset. Cora is the only dataset implemented for now.

Differences with the paper

I'm using a dropout ratio of 0.1 by default (that's the probability of keeping the neurons). With higher values, the models overfit a lot and the results on the validation and test sets are bad.

Cite

If you use this implementation in your research, please cite the paper:

@article{xin2020graph,
  title={Graph Highway Networks},
  author={Xin, Xin and Karatzoglou, Alexandros and Arapakis, Ioannis and Jose, Joemon M},
  journal={arXiv preprint arXiv:2004.04635},
  year={2020}
}