[FIX] Trees: Fix classification from sparse data

[janezd]

Issue

Tree prediction for sparse data does not work.

Description of changes

Added a corresponding function in Cython.

Includes

• Code changes
• Tests

[ajdapretnar]

Works. 🌈 🥇 💫

[nikicc]

Works for me. @janezd, thanks for a prompt fixup!

[janezd]

Oops, this expects csr. @nikicc, it would be trivial to add a function for csc, too. Or do I just call to_csr?

[nikicc]

@janezd I would just call to_csr. Do you have any idea what is going on with the Mosaic test on AppVeyor?

[codecov-io]

Codecov Report

Merging #2496 into master will increase coverage by 0.05%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #2496      +/-   ##
==========================================
+ Coverage   74.53%   74.59%   +0.05%     
==========================================
  Files         321      321              
  Lines       56134    56141       +7     
==========================================
+ Hits        41838    41876      +38     
+ Misses      14296    14265      -31

[janezd]

I added a separate function for csr; it's almost the same, just with different indexing.

I've no idea about Mosaic. Let's see if it fails again.

[kernc]

It's tremendous satisfaction working in a team that agrees on what to optimize for.

>>> import scipy.sparse
>>> csr = scipy.sparse.random(1e5, 1e4, format='csr')
>>> %timeit csr.tocsc()
417 ms ± 2.65 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

[janezd]

Half a second. If the code that I added is executed ten (OK, twenty) times (on the data of this size), the time saved by adding the code will roughly equal the time I spent by adding it.

[kernc]

Thanks, that's exactly what I meant. 😃

[nikicc]

Related to #2370.