Variance values for cell-level uncertainty. about mimir HOT 8 CLOSED

Basic implementation of this is done. I need to test it further. Right now it works as follows,
SELECT VAR(B) FROM SANE_R;
gives the following result

+-------------+
| EXPR_1_VAR |
+-------------+
| 0 |
| 0 |
|8.96187201351742 |
| 0 |
| 0 |
| 0 |
| 0 |
+-------------+

from mimir.

8.96 seems ... a bit high. If memory serves, SANE_R's B attribute can only take values from 1-5. A 9 shouldn't even be possible, even with an even split of samples with B=1 and B= 5.

from mimir.

Just tested it out. The numbers seem a bit more sane. We should do a code review, but otherwise looks good. Let's integrate it into the UI

from mimir.

I had made a mistake with the indexes in the earlier one. But the variance calculation is still wrong. When I get classes from the classifiers getVotesForInstance method, I thought that we get the PDF for the model, but the probabilities don't add up to one. I'm not sure if I have understood something wrong here.

from mimir.

Ok, the returned values were probabilities which the classifiers belonging to each class assign for the given instance. It is not the PDF.
There was a function for variance in the moa Instances class, which when used with the class index returned the required attributes variance. The variance seems to be working correctly now.

from mimir.

Yeah, the only lens we have at the moment is a categorical version of the DCR lens. Realistically, this is not something for which variance will be particularly helpful. We probably need a few more lenses before we can test this process end-to-end rather than with just unit test cases.

from mimir.

Variance is now calculated by sampling.

from mimir.

Cool! Close the ticket after the code review.

from mimir.