RecalibrationFunction

RecalibrationFunction

is an option for Classify, Predict and related functions that specifies how to post-process model predictions.

Details

  • For classifiers, recalibration is also known as probability calibration and is typically used to correct overconfident or underconfident classifiers.
  • RecalibrationFunction can be used at training time, inference time or to update the calibrator of an existing model.
  • When specified at inference time or to update the calibrator of an existing model, typical settings for RecalibrationFunction include:
  • Noneremove existing recalibration
    fappend the function f to the existing calibrator
  • For classifiers, the function f is applied to the class probabilities f[<|class1p1,class2p2,|>] and should return new class probabilities. New class probabilities are automatically normalized.
  • For predictors, the function f transforms the predictive distribution by being applied to the output values.
  • When specified at training time, typical settings for RecalibrationFunction include:
  • Noneprevent any recalibration
    Automaticrecalibrate the model when needed
    Allalways recalibrate the model
  • Besides being used on the final model, recalibration is also applied to candidate models generated by Classify and Predict during their training procedure.
  • When RecalibrationFunctionAll, recalibration is applied to every candidate model.
  • When RecalibrationFunctionAutomatic, recalibration is only used for candidate models that need it.

Examples

open allclose all

Basic Examples  (4)

Load the MNIST dataset:

Train a random forest classifier without any recalibration:

Visualize the calibration curve on a test set:

Train a random forest classifier with recalibration:

Visualize the calibration curve on a test set:

Train a classifier:

Compute the class probabilities of a new example:

Check if the model has been calibrated:

Temporarily set a recalibration function to apply to the probabilities:

Set a permanent recalibration function to apply to the probabilities:

Compute the class probabilities of a new example:

Remove the recalibration function from the classifier:

Load the Boston Homes dataset:

Train a predictor with model calibration:

Visualize the comparison plot on a test set:

Remove the recalibration function from the predictor:

Visualize the new comparison plot:

Train a predictor:

Compute the prediction:

Compute the predictive distribution:

Temporarily set a recalibration function to apply to the prediction:

Compute the predictive distribution with this new recalibration:

Applications  (1)

Load the Titanic dataset:

Train a nearest neighbors classifier with no calibration function:

The classifier is slightly overconfident:

Select the worst classification case in the test set:

Evaluate the estimated probabilities:

Use "temperature scaling" to reduce the classifier self-confidence:

Wolfram Research (2021), RecalibrationFunction, Wolfram Language function, https://reference.wolfram.com/language/ref/RecalibrationFunction.html.

Text

Wolfram Research (2021), RecalibrationFunction, Wolfram Language function, https://reference.wolfram.com/language/ref/RecalibrationFunction.html.

CMS

Wolfram Language. 2021. "RecalibrationFunction." Wolfram Language & System Documentation Center. Wolfram Research. https://reference.wolfram.com/language/ref/RecalibrationFunction.html.

APA

Wolfram Language. (2021). RecalibrationFunction. Wolfram Language & System Documentation Center. Retrieved from https://reference.wolfram.com/language/ref/RecalibrationFunction.html

BibTeX

@misc{reference.wolfram_2024_recalibrationfunction, author="Wolfram Research", title="{RecalibrationFunction}", year="2021", howpublished="\url{https://reference.wolfram.com/language/ref/RecalibrationFunction.html}", note=[Accessed: 15-October-2024 ]}

BibLaTeX

@online{reference.wolfram_2024_recalibrationfunction, organization={Wolfram Research}, title={RecalibrationFunction}, year={2021}, url={https://reference.wolfram.com/language/ref/RecalibrationFunction.html}, note=[Accessed: 15-October-2024 ]}