I’m currently trying to understand if there are Deep Learning techniques based on GNNs that can reconstruct partial areas of a graph. For example, I have a graph with 5 nodes and 10 edges, I know the numerical value of the features for 3 nodes and for all 10 edges, and I would like to build a network capable of making a prediction about the hypothetical value that should appear in the features of the other 2 missing nodes. I was trying to imagine an architecture of this kind, but I find it difficult to visualize how a network based on a GCN or Message Passing could behave with nodes where I do not have a numerical value but where I want to find the optimal one.

I have currently tried consulting the main DL approaches to graphs, but the problem seems to be that this type of task is not addressed. For this reason, I am questioning its feasibility.
( https://distill.pub/2021/gnn-intro/, and some tutorials https://pytorch-geometric.readthedocs.io/en/latest/get_started/colabs.html )