Advice on inconsistent graph sizes?

[JackMaguire]

Hi @danielegrattarola ,

I just found this library and am excited to start using it! I really appreciate the examples and great documentation. After reading, I still have one question: do you have any advice on modeling dynamically sized graphs (where the node count varies from graph to graph)?

My naive instinct is to train a network with a large N and zero-pad the unused elements when the number of nodes is < N. I am not sure if the zeros should be at the beginning, end, or just anywhere. I am interested to hear your input.

Thanks in advance for any advice!
-Jack

[danielegrattarola]

Hi,

there are a couple of solutions that you can try.

If the graphs are small enough to be zero-padded, then you can use spektral.utils.numpy_to_batch to do exactly what you suggest (this is called "batch mode" in the documentation). Note that this will waste a lot of memory because you will have dense arrays with a lot of zeros.

Another possibility is to use "disjoint mode", where you take the disjoint union of your graphs as if they were a single graph. This saves more memory because the adjacency matrix can be a sparse tensor, and also avoids the zero-padding.
Use spektral.utils.numpy_to_disjoint to convert your graphs to the single graph.

Hope this is what you were looking for.

Cheers,
Daniele

[JackMaguire]

Thanks for the quick feedback, Daniele! I will try out your suggestions.

Best,
-Jack