Fold-in an unseen sentence or (doc) from the built graph to obtain its representation (GCN embedding)

[40347015S]

Hi Louis:
Thanks for your contribution. May I ask you a little question?
Instead of using the whole corpus that includes the train and test set for building the graph, for real-world scenarios, it seems natural to "fold-in" a new (unseen) sentence (or doc) that did not appear in the training process from the built graph, just like the fold-in method in the LSA (Latent Semantic Analysis).
With your code, how can I do or modify the code to realize this concept? (Modify the X, A, or W in equation 7 in the paper?)

Thanks!