Observation-wise losses for classification tasks

[jemus42]

Generally, both Brier score and logloss should have observation-wise losses available but implementations in mlr3measures don't include them yet.

Unfortunately they're a bit more involved than adding a single line for a mapping like obs_loss = "se", see mlr-org/mlr3measures#57

As of now I think the only classif measure with an obs loss available is CE / Acc, which is better than nothing but when we want to showcase how we can do CPI for example, it would be nice to have at least logloss availaible.

[Rfa-500]

Hi @jemus42

You should add observation-wise loss functions to mlr3measures for logloss and Brier and wire them to the Measure$obs_loss field, mirroring how zero_one already exposes per-observation errors. That unlocks CPI and similar plots without changing APIs: rr$obs_loss(msr("classif.logloss")) would then return a column of per-row losses.

Useful refs to mirror/plug into:

Measure$obs_loss contract in mlr3 (how per-row losses are surfaced): https://mlr3.mlr-org.com/reference/Measure.html
[mlr3](https://mlr3.mlr-org.com/reference/Measure.html?
Existing logloss measure in mlr3measures (add the obs_loss sibling here): https://mlr3measures.mlr-org.com/reference/logloss.html
mlr3 Measures
Per-observation example (zero_one) showing the expected behavior/shape: https://www.rdocumentation.org/packages/mlr3measures/versions/1.0.0/topics/zero_one
RDocumentation
Brier score definition for the per-row formula you’ll implement: https://en.wikipedia.org/wiki/Brier_score
Wikipedia

This is a small PR: implement obs_loss for both measures (return a numeric vector per row), keep the existing aggregator as is, and the rest of the tooling (e.g., ResampleResult$obs_loss) will work out of the box.