MNT: sklearn deprecation of alpha parameter to NMF

[scottclowe]

In version 1.0.0, sklearn deprecated the alpha argument to NMF in favour of the alpha_W and alpha_H arguments, which were added in v1.0.0. The deprecation warning declares that alpha will be removed in v1.2.0.

We can't simply swap to using alpha_W instead of alpha because the behaviour is different:

When using alpha instead of alpha_W and alpha_H, the regularization terms are not scaled by the n_features (resp. n_samples) factors for W (resp. H).

I imagine the new method where the alphas scale against the number of features/samples is better... but we selected alpha=0.1 without any such scaling and it is not clear to me without running some tests that we should use alpha_W=alpha_H=0.1.

Another issue is that v1.0.0 which added alpha_W and alpha_H only support python >= 3.7, so we would have to cut support for python <= 3.6 to move to the new versions of alpha.

Until someone investigates a new appropriate value of alpha, for the meantime we will carry on using the old version of alpha, suppress the deprecation warning so that end-users do not have to tolerate it, and declare that FISSA requires sklearn<1.2.

[codecov[bot]]

Codecov Report

Merging #271 (de882ba) into master (6b3b320) will increase coverage by 0.01%. The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #271      +/-   ##
==========================================
+ Coverage   94.29%   94.30%   +0.01%     
==========================================
  Files           8        8              
  Lines        1209     1212       +3     
  Branches      310      311       +1     
==========================================
+ Hits         1140     1143       +3     
  Misses         35       35              
  Partials       34       34              

Flag	Coverage Δ
nbsmoke	60.31% <100.00%> (+0.09%)	:arrow_up:
unittests	93.97% <100.00%> (+0.01%)	:arrow_up:

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
fissa/neuropil.py	95.38% <100.00%> (+0.22%)	:arrow_up:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 6b3b320...de882ba. Read the comment docs.