Link Prediction on Unseen Graphs

As I’m new to GNNs, I’d just like to check the approach is viable.

I have a set of training graphs, each with a different number of nodes which represent clusters of particles. Each has features like position, velocity, mass etc. They are taken at different timesteps from a number of simulations.

They can be connected by edges, labelled “Will Merge” in the next timestep.

They will additionally be connected by a dummy edge for message passing, as only ~1% of nodes will merge by the next timestep.

I want to train on this set of graphs, then present totally unseen graphs and have a model predict the “Will Merge” edges. I.e., inductive learning. These graphs will have a different set of nodes than has been seen before, and will have no “Will Merge” edges present.

Is this possible? Can you point to any examples of fully inductive link prediction on unseen graphs?
Examples on the docs mostly seem transductive or semi-inductive, but I may just be misunderstanding them as a beginner.

I found this post Inductive Link Prediction on Heterogenous Graphs but no more concrete examples. It seems that post resolves to include all training graphs in a single disconnected supergraph.

And is there a better approach than this?

As long as you train the same model on all these graphs, then generalizing to other graphs should be natural.

One simple way to do so is to have a nested loop for training, the outer loop cycles over different graphs, and the inner loop cycles over the different edges in the same graph like any link prediction on a single graph does.