[ENH] PCA: Output variance of components

[janezd]

Issue

Fixes two thirds of #5468.

To test before biolab/orange-widget-base#162 is merged, change

        self.varmodel[:] = zip(self._variance_ratio[:self.maxp],
                               numpy.cumsum(self._variance_ratio))

to

        self.varmodel.wrap(list(zip(self._variance_ratio[:self.maxp],
                               numpy.cumsum(self._variance_ratio))))

Description of changes

1. Variances are added as attributes of variables, and as extra meta columns in the components output.

2. The table with variances is shown in control area and can be used as the third mean to select components. Clicking a row selects all components up to that row. (Selecting arbitrary rows is not possible: the widget's code does not support it; interactions the other two controls would be weird; this is not typical for PCA; for atypical uses, there's Select Rows; no a/b testing of this feature will be done, thanks for not asking :).

   

3. The third point in the issue was reporting variance on new data. This can be added to the table, but I'd suggest this to be a separate issue/PR.

Includes

• Code changes
• Tests
• Documentation

[codecov]

Codecov Report

Merging #5513 (3cd179c) into master (f9d78d6) will decrease coverage by 0.01%.
The diff coverage is 79.31%.

❗ Current head 3cd179c differs from pull request most recent head a2afed5. Consider uploading reports for the commit a2afed5 to get more accurate results

@@            Coverage Diff             @@
##           master    #5513      +/-   ##
==========================================
- Coverage   86.38%   86.36%   -0.02%     
==========================================
  Files         304      304              
  Lines       61753    61848      +95     
==========================================
+ Hits        53344    53414      +70     
- Misses       8409     8434      +25     

[janezd]

I added variance on test data. (Not thoroughly checked, yet.) The table is a bit wide, but OK. The plot may look nice

... except when it doesn't:

I hate to think about how to compute the positions of labels in pyqtgraph to prevent overlapping. Any good suggestions?

[ajdapretnar]

I love how this widget got from basic to jacked 💪 in no time. 😆

As for the label, if points for labeling are <= to the height of the font, then plot the higher one above, not below the point it labels? (stupid suggestion probably, just throwing ideas out there)

[janezd]

The problem is that in pyqtgraph you set the position in plot coordinates, not in pixels. There are several functions that transform coordinates it's all black magic, at least to @VesnaT and me. I'm thus not even sure I can detect that labels are overlapping.

Ha! Labels are QGraphicsTextItem's, so Qt tells you if they collide. In case they do, I just set the anchors as you suggested and it works like magic. Sometimes the label would overlap with the curve, which didn't happen before because curves were monotone. But it was good for my grandmother, Lord it's good enough for me!

[markotoplak]

I rebased this and am (for now) only merging the first commit. The others are backed-up at https://github.com/markotoplak/orange3/tree/pca-show-variance-backup

In between, a group from Mannheim, particularly @Les-Simon, proposed some changes to the PCA (and other widgets): https://github.com/Charisma-Mannheim/OrangeExtension. They tend to work with PCA a lot, much more than we do. We should consider their suggestions before adding additional complexity to the widget.