Layouts meta-issue

This issue is to organize some thoughts on changes I’d like to see for our layout code. Happy to have feedback.

Is your feature request related to a problem? Please describe.

• Layouts should run on directed networks (see #655). If the user wants to color by automatically detected community, then I’d suggest that we just symmetrize the graph internally (only for the community detection step) and proceed. I don’t think this is overstepping our bounds as to what the user wanted, personally, because this is typically the notion of community for directed graphs.
• Automatic layouts should also be possible using ASE/LSE. I have had a lot of success generating layouts that I like using ASE as opposed to n2v so I think “algorithm” could just be a parameter.
• Allow passing dictionaries of arguments to stage 1 (n2v/ASE/LSE) and stage 2 (umap/tsne) (see #612)
• Should have an interface to the layouts code that feels like what we have in .plot and seaborn. I think we could refactor the code such that we can still have all-in-one functions that eat graphs and spit out plots like layouts.save_graph, but there’s no reason we can’t also let people interact with some of these functions in a way that’s more like matplotlib/seaborn.
• NoOverlap should be an option the user can turn on or off.
  • Addressed in #758
• Layouts should run on adjacency matrices also

Provide references (if applicable)

I have some code here which is closer to the flavor of what I’d want to see for graph plotting code that works with seaborn/matplotlib. Just a reference, not necessarily how we should do anything here. One thing I never figured out for that code which would be nice is automatic scaling of edge/node widths/sizes etc. based on the size of the graph, so that an initial shot at an embedding at least looks reasonable.

[daxpryce]

@bryantower is on spring break and I’m about to check out for the week too, but a quick check of the docs says that you may want a larger perplexity for bigger graphs (though I’m not sure that particular sbm really qualifies)

That said, you should expect this to take a fair amount of time - especially with the no overlap portion on!

[daxpryce]

Do you think we need to warn the user that this is happening or just do it automatically and explain in the docs?

We probably should warn. I don’t like automatically inferring/doing things without warning someone - just in case our “we think this is probably safe” ends up being not safe, but the user won’t ever know we’re trying to be too clever for our own good