Bugs in Metrics

[Roffild]

gamma-nloglik:

bst_float psi = 1.0;
bst_float c = 1. / psi * std::log(y/psi) - std::log(y) - common::LogGamma(1. / psi);

c == 0

logloss:

  XGBOOST_DEVICE bst_float EvalRow(bst_float y, bst_float py) const {
    const bst_float eps = 1e-16f;
    const bst_float pneg = 1.0f - py;
    if (py < eps) {
      return -y * std::log(eps) - (1.0f - y)  * std::log(1.0f - eps);
    } else if (pneg < eps) {
      return -y * std::log(1.0f - eps) - (1.0f - y)  * std::log(eps);
    } else {
      return -y * std::log(py) - (1.0f - y) * std::log(pneg);
    }
  }

std::log(1.0f - eps) == std::log(1.0) == 0

gamma-deviance needs to be removed because the formula is not correct!(#6728)

I don't understand what math formula was used in poisson-nloglik.
-log( Poisson_regression )

With tweedie-nloglik it is also unclear. And the test is missing.

Tests for regression metrics with weights.(#6729)

If metrics are used in forest creation.....

[Roffild]

I wanted to use metrics from XGBoost for Pytorch.

But now only sklearn.metrics for all models!

[trivialfis]

I need to take a closer look.

[trivialfis]

The bug in gamma deviance is fixed. Better documentation for other metrics will be a different topic. Thanks for raising the issue!

[trivialfis]

Just a quick note for everyone who has been following this thread. I believe these metrics and objectives are derived from the generalized linear model.

[Roffild]

All metrics are calculated for each result separately.

Loss is calculated for each individual result, but metrics must be calculated for the entire result matrix. Therefore, the metrics in XGBoost are approximate.

[Roffild]

pytorch/pytorch#22439

[trivialfis]

I don't understand what math formula was used in poisson-nloglik.

See Evaluating the Poisson distribution section of https://en.wikipedia.org/wiki/Poisson_distribution

[trivialfis]

gamma-nloglik:

Yeah, this one is a bit confusing, I tracked down the PR for -log(\gamma): #1369 , which hard coded the dispersion to 1. Not entirely sure why.

Original PR for adding gamma regression: #1258 .

[trivialfis]

The weird logloss you see is just a way to work around numerical issues.