A calibrated classifier is one where the predicted probability of a class closely matches the rate at which that class occurs, e.g. for data points which are assigned a predicted probability of class A of .8, approximately 80 percent of such points should belong to class A if the classifier is well calibrated. This is estimated empirically by grouping data points with similar predicted probabilities for each class, and plotting the rate of each class within each bin against the predicted probability bins.

generateCalibrationData(obj, breaks = "Sturges", groups = NULL,
  task.id = NULL)

Arguments

obj

(list of Prediction | list of ResampleResult | BenchmarkResult)
Single prediction object, list of them, single resample result, list of them, or a benchmark result. In case of a list probably produced by different learners you want to compare, then name the list with the names you want to see in the plots, probably learner shortnames or ids.

breaks

(character(1) | numeric)
If character(1), the algorithm to use in generating probability bins. See hist for details. If numeric, the cut points for the bins. Default is “Sturges”.

groups

(integer(1))
The number of bins to construct. If specified, breaks is ignored. Default is NULL.

task.id

(character(1))
Selected task in BenchmarkResult to do plots for, ignored otherwise. Default is first task.

Value

CalibrationData. A list containing:

proportion

data.frame with columns:

  • Learner Name of learner.

  • bin Bins calculated according to the breaks or groups argument.

  • Class Class labels (for binary classification only the positive class).

  • Proportion Proportion of observations from class Class among all observations with posterior probabilities of class Class within the interval given in bin.

data

data.frame with columns:

  • Learner Name of learner.

  • truth True class label.

  • Class Class labels (for binary classification only the positive class).

  • Probability Predicted posterior probability of Class.

  • bin Bin corresponding to Probability.

task

(TaskDesc)
Task description.

References

Vuk, Miha, and Curk, Tomaz. “ROC Curve, Lift Chart, and Calibration Plot.” Metodoloski zvezki. Vol. 3. No. 1 (2006): 89-108.

See also