precrec 0.16.0

This release is on GitHub only. CRAN still carries 0.14.5, and this version
is not being submitted there yet.

remotes::install_github("evalclass/precrec")

It also carries the 0.15.0 changes, which were prepared but never released;
see NEWS.md for those.

Changes in 0.16.0

• Add metric_curve(), which takes the name of a measure for the x axis and
  the name of a measure for the y axis and draws one against the other, the
  way ROCR::performance() does. Every measure evalmod() can calculate is
  available on both axes, under its own name or under the ROCR identifier.
  It returns an ssxycurves, msxycurves, smxycurves or mmxycurves
  object, chosen the way evalmod() chooses between its own, and the object
  works with print(), as.data.frame(), fortify(), plot() and
  autoplot().

  Only two pairs are joined by a line: false positive rate against
  sensitivity, which is the ROC curve, and sensitivity against precision,
  which is the precision-recall curve. Those two have a defined
  interpolation, and for them metric_curve() hands the work to the same
  code evalmod(mode = "rocprc") uses, so the two cannot disagree. Every
  other pair is drawn as points, because joining raw per-cutoff points with
  straight lines is the error this package was written to avoid; pass
  type = "l" to join them anyway.

  metric_curve() draws one curve per test dataset and does not average
  over them. An average needs a rule for interpolating between the points of
  each curve, which is what an unregistered pair does not have.

• Add the evaluation measures ROCR provides that precrec did not:
  fpr, fnr, false_discovery_rate, false_omission_rate,
  predicted_positive_rate, predicted_negative_rate, lift, odds,
  mi, chisq and cost. Each also answers to the identifier ROCR uses
  for it - fall, miss, pcfall, pcmiss, rpp, rnp,
  mutual_information - and to its standard abbreviation where it has one,
  so a call written against ROCR keeps working. evalmod() gains
  cost_fp and cost_fn for the two weights the cost measure takes;
  with the default weights of 1 it is the error rate.

  They are not calculated unless asked for. evalmod() gains a metrics
  argument that names the measures to add, or takes "all"; the default
  NULL is the fourteen measures the function has always returned, so an
  existing call gets the same object with the same fourteen plot() and
  autoplot() panels. A measure that was not calculated cannot be plotted,
  and the error says which argument asks for it.

  The odds ratio and the chi-square statistic are NA at the top and the
  bottom of every dataset, where the 2x2 table has an empty cell and neither
  is defined. ROCR reports an infinity or a NaN there. NA is what
  precision and npv already do with their own undefined end, and it
  keeps an infinity off a shared axis. The mutual information is 0 at
  those two points rather than NA, because a cutoff that predicts one
  class for everything carries no information about the labels - that value
  is defined, and it is zero.

• Add prbe(), which finds the points of a precision-recall curve at which
  precision and recall are equal. It takes the object evalmod() returns and
  gives back a data frame with one row per break-even point, in the manner of
  auc().

  ROCR::performance(pred, "prbe") interpolates linearly between adjacent
  raw precision-recall points to find the crossing, which is not correct and
  is the reason this package exists; prbe() reads the crossing off the
  curve evalmod() has already interpolated properly.

• Add the sar measure, the mean of accuracy, the AUC of the ROC curve, and
  one minus the root mean squared error. Like the other added measures it is
  opt-in through evalmod(metrics = ). The RMSE reads the values of the
  scores rather than their ranks, so sar warns and returns NA when the
  scores are not probabilities between 0 and 1; every other measure asked for
  in the same call is still returned.

• Add the root mean squared error to prob_metrics() and
  prob_metrics_ci(), as the "rmse" metric. It is the square root of the
  Brier score, so each model and dataset now takes up three rows rather than
  two.

• Declare stats in Imports. It was used but not listed.

• Fix the y axis of plot() for informedness and markedness. Both run from
  -1 to 1, and both were drawn on a 0 to 1 axis, which cut off the negative
  half of the curve. autoplot() was never affected. The axis range of a
  measure now comes from one table rather than from a list of names that had
  the internal short names missing from it.