[FIX] Statistics.util.stats: Fix negative #nans for sparse

[nikicc]

Computing the number of nans for sparse matrices was broken and sometimes returned negative numbers.

Variable non_zero contains the number of defined values in each column. So to calculate the number of undefined (nans) for each column we have to substract non_zero from the number of values in the column i.e. X.shape[0] and not X.shape[1]. The bug was noticed when the number of defined values in some column was larger than the number of features and consequenlty the number of nans was negative. This caused density in OWTable to exceeded 100%.

[codecov-io]

Current coverage is 88.73% (diff: 100%)

Merging #1659 into master will not change coverage

@@             master      #1659   diff @@
==========================================
  Files            78         78          
  Lines          8150       8150          
  Methods           0          0          
  Messages          0          0          
  Branches          0          0          
==========================================
  Hits           7232       7232          
  Misses          918        918          
  Partials          0          0          

Powered by Codecov. Last update 282b1b0...7b35830

[lanzagar]

If the wrong computation was unnoticed, that probably means there is no test for nan counting. Could you add a simple one?

[nikicc]

@lanzagar IMHO we already have a test for this — the one that I corrected in this PR — just that it was incorrect.

For sparse data "nans" are actually values not present in sparse matrix and the test_stats_sparse already contain those since csr_matrix does not store zeros.

[lanzagar]

You are right, the (not very clear) test that was already there (but broken) does test this too. I guess it is enough to cover the basics now that it is fixed.