ENH: add subgraph method to Graph to get subsets

[martinfleis]

I have found myself in a need to create subsets of a large graph covering only specific portions of my geometries. Hence I have developed a method to create a subgraph since making a subset of adjacency and passing that to a constructor discards isolates.

[martinfleis]

This turned out to be several orders of magnitude faster for large string indices.

[codecov]

Codecov Report

Merging #640 (77bd0ff) into main (79c4f82) will increase coverage by 0.0%.
Report is 1 commits behind head on main.
The diff coverage is 100.0%.

@@          Coverage Diff          @@
##            main    #640   +/-   ##
=====================================
  Coverage   83.9%   83.9%           
=====================================
  Files        139     139           
  Lines      14970   14976    +6     
=====================================
+ Hits       12562   12569    +7     
+ Misses      2408    2407    -1     

Files	Coverage Δ
libpysal/graph/_utils.py	89.3% <100.0%> (ø)
libpysal/graph/base.py	97.7% <100.0%> (+<0.1%)	⬆️
libpysal/graph/tests/test_base.py	100.0% <100.0%> (ø)

... and 6 files with indirect coverage changes

[knaaptime]

cool. I wonder if we can use this to enhance pysal/esda#259 as discussed over there (e.g by computing the largest range first, then successively cutting down the graph). Seems like you'd still need a tree query to get the indices though?

[martinfleis]

I wonder if we can use this to enhance pysal/esda#259 as discussed over there (e.g by computing the largest range first, then successively cutting down the graph). Seems like you'd still need a tree query to get the indices though?

I don't think so. This aims to create a subgraph based on a subset of focals. In correlogram, you need to keep the same focals but cut their neighbors.

What you could doe once #635 is in is something like this:

for i in distances:
    adj = graph_w_distance.adjacency.copy()
    adj[adj > i] = 0
    smaller_graph = graph.Graph(adj, is_sorted=True).transform("r")
    # compute stats using smaller_graph

This assumes that graph_w_distance has weight == distance. Right now, there is the sorting bottleneck in this code but after #635 that will be gone.

edit: you could eventually call eliminate_zeros() from #634 on that as well before transform to get a cleaner graph but I am less sure about perf benefits of that.