[WIP] Node reductions - #283

@lmeyerov Despite this being pretty much a work in progress, I already opened a PR to get some initial feedback. I hope that's okay :)

This is a partial implementation of table reductions. The function replace_nodes_with_edges(node_selector) can currently accept both a string and a list of strings. For example the following calls g.replace_nodes_with_edges(['m']) and g.replace_nodes_with_edges('m') both yield the same results. Here is an example:

g = graphistry.edges(pd.DataFrame([{'x': 'a', 'y': 'm'},
                                   {'x': 'b', 'y': 'm'},
                                   {'x': 'c', 'y': 'n'},
                                   {'x': 'd', 'y': 'm'}]), 'x', 'y')
g.replace_nodes_with_edges(['m',])
print(g._edges)

#result
   x  y reduced_from
0  a  b            m
1  a  d            m
2  b  d            m
3  c  n          NaN

What are your thoughts?

Cool, some thoughts:

• what happens if m is a src (x), not dst (y)?
• how about specify nodes as a Series, say 1-column or if multi-column, index on g._node, to help composition?

Ex:

annoying_nodes_df = g._nodes[ g._nodes['type'] == 'event' ]
g2 = g.replace_nodes_with_edges(annoying_nodes_df)

• reduced_from right now (afaict) would report "m,n,o,p" instead of "m" when used on a list input

I'll need to look at itertools, but the rest seems straightforward to port to cudf / dask_cudf :)

@lmeyerov Thank you very much for your feedback. I have pushed another commit now. Kindly have a look at the replace_nodes_with_edges(..) function comments, it should address some of the issues your raised. I have also added some tests to show how to use this function. As far itertools, I have replaced the logic which now uses numpy.tril. Please have a look and let me know what you think. :)

For now I would like to make some headway into getting the table reductions into a good shape before starting to work on pregel/graphx style mapping.

Great!

In case it helped, I sketched out a few the top use cases we see in the original ticket: #193

re:tril, I'm struggling to understand that logic, which suggests someone doing a future bugfix here would need to rewrite it.. can it be something like pd.concat? I'm guessing once the aggregation operators come in and more tests, we'll have a better sense of what needs to be fast/slow, so simple vector pd calls should get us far..

Just checking in on this. Maybe the thing to do is:

• make sure each of the motivating examples is in a test
• test a few different tricky graph shapes

I'm thinking we can land this as-is, and update the underlying feature request for future PRs to explore (a) a custom reductions mode for this (ex: summary stats) and (b) multi-node collapsing/grouping (ex: collapse connected nodes name/email/address/username into a single labeled node)