I want to train a model for link prediction. The input would be one central graph + a set of smaller graphs, and I want to predict links between each smaller graphs’s nodes and the central graph’s nodes. However it is unclear to me how the graphs should be input to my model since disconnected graphs are used for batch processing. From what I understand a batch of graphs is meant to be processed as several separate instances of the same problem, not one instance made of several disconnected graphs.

Is there a recommended approach for such problems ? One idea I had was to use a heterograph instead, where each small graph and the central graph would have one dedicated node type. Additionally I would add one dummy node connected to every other node to make my heterograph connected.

Am I missing something or is it the right approach ?

Thanks in advance for the help !