Imagine you have a set of samples for which you can represent the data as graphs. Each graph is not at all the same size or shape as any of the others.

When trying to use a MPNN for graph classification, I run into what seems to be a really unsatisfying predicament.

After performing the message passing steps, because I have a differential number of vectors across samples (because the graphs are different sizes), I'm forced to just average all of the new embeddings to create a singular standard vector to carry through.

I feel like this leads to dramatic information loss. What was the point of representing the whole system and seeing how neighbors can influence each other only to squash everything in the end?

Does anyone know a better way?