OWNomogram: Add a new widget

[VesnaT]

Issue

Description of changes

Includes

• Code changes
• Tests
• Documentation

[codecov-io]

Current coverage is 89.51% (diff: 100%)

Merging #1936 into master will decrease coverage by 0.01%

@@             master      #1936   diff @@
==========================================
  Files            90         90          
  Lines          9180       9182     +2   
  Methods           0          0          
  Messages          0          0          
  Branches          0          0          
==========================================
  Hits           8219       8219          
- Misses          961        963     +2   
  Partials          0          0          

Powered by Codecov. Last update f3ebda7...5f781ce

[lanzagar]

A couple of suggestions:

1. Some difference in the font for Points/Total/Probabilities vs. feature names would be very welcome. I think just changing feature names to not be bold would already make the visualization easier and quicker to process.
2. Log. Reg. models always use align left (setting disabled). And for Naive Bayes, center alignment will mostly be used. I suggest removing this setting completely, and just selecting this automatically based on the classifier.
3. I suggest removing the option Variance in Sort by. Together with absolute importance they are a bit redundant, as the latter incorporates it to a degree anyway - a variable can't distinguish classes well without enough variance. Variance of a variable is also not readable from this visualization (unlike other measures) and should be investigated somewhere else (box plot?).
4. A better default would probably be to show the top 10 most informative variables (absolute importance). All and unsorted is probably going to be changed very quickly and every time for data with tens of variables or more.
5. I was not yet able to figure out what the Continuous features: option is used for? It is disabled for data with discrete and/or continuous features.
6. Total and Probabilities appear as separate graphical elements. IMHO it would be better to either:
   a) Have them actually be separate. In this case Probabilities could be on a linear scale from 0 to 1. (points not aligned)
   b) Use one line and one point, with values for Total/Prob. shown above/below.
   c) Keep two separate scales, but make it more clear they are connected by a small vertical line going through both (aligned) points.
   With b) and c) users are less likely to look only at Prob. (while changing feature sliders) and be confused by some peculiarities (like the cutoff and a deformed scale). Adding <= and >= to the first and last value could also help.

[VesnaT]

@janezd Could you please confirm the above suggestions?

[janezd]

@VesnaT, great. Snapping to the closest value is a cherry on top.

@lanzagar, I do have a few more ideas ... but if you agree, we'd merge this now and improve later.

[lanzagar]

We noticed some bugs with @VesnaT yesterday, which she began fixing. When she confirms that is finished I am in favour of merging, and doing improvements in new PRs afterwards.