ENH: add greedy (topological) coloring

[martinfleis]

Implementation of greedy package for topological (greedy) colouring of GeoPandas GeoDataFrames. This is the first draft to have something concrete to work on. It is essentially copy&paste of the original code from https://github.com/martinfleis/greedy with few changes here and there to mention that greedy.greedy is now mapclassify.greedy and to make deps optional. Three notebooks used as docs for greedy are now merged to one to become part of Pysal Notebooks.

Apart from the obvious questions whether this is the optimal way how to implement it within mapclassify, there are some additional which need to be resolved.

1. License. The core of _balanced algorithm is ported from QGIS where it is licensed under GPL-3.0. To make my life easier, I have used the same license for whole greedy. We need to figure out how to resolve this as GPL-3.0 is not compatible with BSD-3-Clause afaik. I don't mind for my code and @nyalldawson mentioned that he's also willing to relicense that part coming from QGIS if needed.

2. Terminology. QGIS uses topological colouring, I am using mostly greedy colouring. What do you prefer? It would be good to have continuity from existing greedy, but I don't see it as a crucial thing if we decide that topological coloring is a preferred way. We can change greedy to something else.

3. _geos_sw is a function which generates libpysal.W object based on intersection between objects or their buffered versions. Naturally it is significantly slower that Queen or Rook weights (but that will get better with pygeos within GeoPandas eventually), but it does not require topological correctness of dataset and allows tolerance (i.e. buffer). I feel that the right place for it could be libpysal as it may be useful elsewhere apart from greedy colouring. Should we implement it here with a note that it should go to libpysal and remove it from here once it is released there?

For details on what greedy brings, see https://martinfleis.github.io/greedy/.

xref pysal/splot#94, geopandas/geopandas#1165

[knaaptime]

sweet, thanks! Others may have opinions here, but my quick responses would be

1. agree this will need to be resolved. we're kinda sticklers for BSD-3, so if you're willing to relicense, that would be the path of least resistance

2. i vote for topological coloring, as I personally find it far more intuitive but i dont have a terribly strong preference

3. how does _geos_sw compare to libpysal.weights.fuzzy_contiguity?

[martinfleis]

@knaaptime I had no idea fuzzy_contiguity exists, to be honest. I see only one difference and that is in the way how tolerance is set. While in fuzzy_contiguity that is a fraction of gdf bbox length, in _geos_sw the actual distance between features. Otherwise the algorithm is essentially the same, just buffer in here is vectorised unlike in libpysal. I guess that the best way forward is proposing minor enhancement to fuzzy_contiguity to allow both ways of setting tolerance.

[martinfleis]

I have marked most of the tests with pytest.xfail as there is no rtree on CI. That would be ideal to get from conda-forge, but apart from numba everything now comes from pip (I tried to get rtree but it resulted in error https://travis-ci.org/github/pysal/mapclassify/builds/673150902). Do you have a preference how to go around that? We may even keep it xfailing if this functionality will eventually come from fuzzy_contiguity.

[knaaptime]

we should just install everything from conda (like we do elsewhere) and avoid pip on CI altogether

[nyalldawson]

Just to confirm that I'm happy to relicense any part of this I've written to whatever license works for your needs. Even WTFPL if you want.

[knaaptime]

@nyalldawson 😁 appreciate it!

[martinfleis]

I have opened a small PR in libpysal pysal/libpysal#280 extending functionality of existing fuzzy_contiguity to accommodate the needs of greedy (specific buffer distance). That resolves the question of _geos_sw which will be replaced by libpysal.fuzzy_contiguity as soon as this update is released.

[sjsrey]

This looks very nice.

I left one question about the license.

[martinfleis]

Thanks @sjsrey. I have removed the explicit license from the file as it will inherit the license of the whole package now.