Node Features Propagation and Prediction

[wijuliano]

Hello everyone, I was wondering if I can use StellarGraph to do a node feature prediction.

I basically have a graph that contain about 1000 nodes and only 100 of them contain a feature that range from 0 to 1. I wanted to create a graph attention network that is able to predict the missing node features. I though I could add a propagation layer where the features of the nodes are dependent on its neighbors and then a linear regression layer at the end to predict the new node features. Is that possible? I am running into so many errors/troubleshooting.

I am also having trouble in splitting my dataset into training and validation. I have been trying to mask my missing features and only compare the loss from the predicted nodes with the ones that I already have data on.

If anyone have a code similar, it would really helpful. All of the codes that I have seen so far use node features to predict node labels.