[ENH] PCA: Preserve f32s & reduce memory footprint when computing means

[pavlin-policar]

Issue

While working towards getting the improved PCA merged into scikit-learn, I've found two improvements.

Description of changes

1. np.float32 are now preserved
2. Apparently, scipys sparse x.mean method isn't the most memory efficient, and scikit-learn's utility function mean_variance_axis is much better (see benchmarks here)

Includes

• Code changes
• Tests
• Documentation

[codecov]

Codecov Report

Merging #3582 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #3582      +/-   ##
==========================================
+ Coverage   83.98%   83.98%   +<.01%     
==========================================
  Files         370      370              
  Lines       66976    66981       +5     
==========================================
+ Hits        56249    56254       +5     
  Misses      10727    10727

[pavlin-policar]

Codecov makes no sense. I added code with no tests, yet still somehow managed to improve coverage. What?

[lanzagar]

We already have Orange.statistics.util.mean which is supposed to handle dense and sparse matrices, but it does not have the axis parameter (unlike most other functions in that module). Maybe you could improve that function instead and use it here?

[pavlin-policar]

That's a much better idea. Unfortunately, mean_variance_axis just ignores NaNs and provides no feedback if the data had any NaNs. This means that mean wouldn't properly handle NaNs and we'd have no good way of knowing where they occurred.

However, changing nanmean to use this is completely fine since this is exactly the behaviour we want there. So I did that.

[lanzagar]

If you add new code with no new tests it can still be covered by existing tests, thus increasing the coverage... ;)