roc.curves: Plot ROC Curves

Description

This function plots ROC Curves of one or several classification predictions.

Usage

roc.curves(
  predictions,
  gt,
  methods.names = NULL,
  positive = levels(factor(gt))[1],
  type = c("auto", "fuzzy", "hard"),
  ...
)

Value

Nothing; the curves are drawn on the current graphics device.

Arguments

predictions: The predictions of one or several classification models. Four shapes are accepted: a factor of hard labels (one model); a numeric vector of scores for the positive class (one model); a matrix of class probabilities, i.e. one column per class named after it, as returned by predict (model, x, fuzzy = TRUE) (one model); or any other matrix/data.frame, read as one column per model.
gt: Actual labels of the dataset (factor or vector), two classes only.
methods.names: The name of the compared methods (vector).
positive: The label of the positive class. Defaults to the first level of gt, as everywhere else in the package -- note that for the usual alphabetical ordering this is often the negative class, so it is worth setting explicitly.
type: "auto" (default) reads predictions according to its shape, "fuzzy" requires scores and refuses hard labels, "hard" reduces everything to the predicted class first (this is the coarse three-point curve discussed above).
...: Other parameters, passed to the underlying plot.

Details

A ROC curve needs a score: the higher it is, the more likely the observation is to belong to the positive class. The natural one is the estimated probability returned by predict (model, x, fuzzy = TRUE). Hard class labels give only two distinct values, so the "curve" reduces to three points and the area under it says very little. That coarse version is available on purpose -- it makes a useful comparison -- but it has to be asked for: pass hard labels, or type = "hard".

Examples

Run this code

require (datasets)
data (iris)
d = iris
levels (d [, 5]) = c ("+", "+", "-") # Building a two classes dataset
model.nb = NB (d [, -5], d [, 5])
model.lda = LDA (d [, -5], d [, 5])
# From the estimated probabilities: the meaningful curve
roc.curves (predict (model.nb, d [, -5], fuzzy = TRUE), d [, 5], positive = "+")
# Two models compared, one score column each
roc.curves (cbind (NB = predict (model.nb, d [, -5], fuzzy = TRUE) [, "+"],
                   LDA = predict (model.lda, d [, -5], fuzzy = TRUE) [, "+"]),
            d [, 5], c ("NB", "LDA"), positive = "+")
# The same predictions reduced to hard labels: three points, and little to read
roc.curves (predict (model.nb, d [, -5]), d [, 5], positive = "+", type = "hard")