How to apply this to graphs?
One Paper (Liu 2019) uses two NN:
Autoencoder graph->vector
NF on vector data
which is fine, but also not really graph specific
No interaction between encoding and transformation