A single embedding of a graph batch

Questions
#1

Hi,

I am doing a graph embedding task using a GCN model (Kipf, 2017). My inputs are a lot of small (<5) batches of graphs. I was wondering how is it better to get a single embedding for each single batch, so that it can used all at once in the downstream regression?

There is also a second part of a question. Those batches belong to a few distinct classes (batches >> classes), there is a ‘one batch: many classes’ mapping. This class information does not need to be propagated through the graph structures, it is a label. How to better add this information?

Right now i am getting embeddings of single graphs from each batch, then concatenating them and then adding labels,.

thanks!

#2
  1. Your scenario sounds like graph regression. Have you checked the user guide on graph classification?
  2. You can project the classes and concatenate them with the learned embeddings of the graphs.
#3

this is brilliant! and I feel silly that i missed. Thank you

I am doing a regression on graph emdeddings. But yea, not much different

have a good day

1 Like