negate the result and prefix the metric name for error/loss metrics

[sebhrusen]

#268

Only changing the result and metric columns when the metric represents an error:

• metric name is prefixed with neg_.
• result is negated.

Example:

Summing up scores for current run:
                  id         task          framework constraint fold    result       metric   mode version params               app_version                  utc  duration  training_duration  predict_duration models_count       seed       acc  auc    balacc   logloss      mae        r2     rmse
0  openml.org/t/3913          kc2  constantpredictor       test    0   0.50000          auc  local  0.23.2         dev [issue/268, b990e93]  2021-04-01T15:52:54       0.2                0.0               0.0            1  933769621  0.792453  0.5  0.500000  0.510714      NaN       NaN      NaN
1  openml.org/t/3913          kc2  constantpredictor       test    1   0.50000          auc  local  0.23.2         dev [issue/268, b990e93]  2021-04-01T15:52:54       0.1                0.0               0.0            1  933769622  0.792453  0.5  0.500000  0.510714      NaN       NaN      NaN
2    openml.org/t/59         iris  constantpredictor       test    0  -1.09861  neg_logloss  local  0.23.2         dev [issue/268, b990e93]  2021-04-01T15:52:54       0.0                0.0               0.0            1  933769621  0.333333  NaN  0.333333  1.098610      NaN       NaN      NaN
3    openml.org/t/59         iris  constantpredictor       test    1  -1.09861  neg_logloss  local  0.23.2         dev [issue/268, b990e93]  2021-04-01T15:52:54       0.0                0.0               0.0            1  933769622  0.333333  NaN  0.333333  1.098610      NaN       NaN      NaN
4  openml.org/t/2295  cholesterol  constantpredictor       test    0 -45.68970     neg_rmse  local  0.23.2         dev [issue/268, b990e93]  2021-04-01T15:52:54       0.0                0.0               0.0            1  933769621       NaN  NaN       NaN       NaN  35.6774 -0.077562  45.6897
5  openml.org/t/2295  cholesterol  constantpredictor       test    1 -55.00410     neg_rmse  local  0.23.2         dev [issue/268, b990e93]  2021-04-01T15:52:54       0.0                0.0               0.0            1  933769622       NaN  NaN       NaN       NaN  45.3871 -0.049133  55.0041

[sebhrusen]

@Innixma does this look reasonable?

[PGijsbers]

Prefixing neg_ is probably more sensible (since scikit-learn does it) so I agree with that choice even though it takes more space. Other than that I tested it (with constant predictor) and it looks good.

[Innixma]

Apologies for late response, I was on PTO.

Looks good to me, had a comment for long-term improving code quality and extensibility.

[Innixma]

nit: area instead of array

[sebhrusen]

oups! will fix

[Innixma]

This seems a bit hacky. Better to have either a dictionary mapping or metrics as classes (example in AutoGluon).

[sebhrusen]

I can't disagree with you: it IS a bit hacky.
Ideally, there should be a class for each metric. It's probably something I'll do at some point to support custom metrics or other customizations in a more satisfying way that what was done in #141.
If there's a demand for it, I'll do it.