tic Lifecycle: experimental CRAN Status DOI License: GPL v3 Mattermost

What does it do?

mcboost implements Multi-Calibration Boosting (Hebert-Johnson et al., 2018; Kim et al., 2019) for the multi-calibration of a machine learning model’s prediction. Multi-Calibration works best in scenarios where the underlying data & labels are unbiased but a bias is introduced within the algorithm’s fitting procedure. This is often the case, e.g. when an algorithm fits a majority population while ignoring or under-fitting minority populations.

For more information and example, see the package’s website.

More details with respect to usage and the procedures can be found in the package vignettes.


The current version can be downloaded from CRAN using:


You can install the development version of mcboost from Github with:



Post-processing with mcboost needs three components. We start with an initial prediction model (1) and an auditing algorithm (2) that may be customized by the user. The auditing algorithm then runs Multi-Calibration-Boosting on a labeled auditing dataset (3). The resulting model can be used for obtaining multi-calibrated predictions.


In this simple example, our goal is to improve calibration for an initial predictor, e.g. a ML algorithm trained on an initial task. Internally, mcboost often makes use of mlr3 and learners that come with mlr3learners.


First we set up an example dataset.

  #  Example Data: Sonar Task
  tsk = tsk("sonar")
  tid = sample(tsk$row_ids, 100) # 100 rows for training
  train_data = tsk$data(cols = tsk$feature_names, rows = tid)
  train_labels = tsk$data(cols = tsk$target_names, rows = tid)[[1]]

To provide an example, we assume that we have already a learner l which we train below. We can now wrap this initial learner’s predict function for use with mcboost, since mcboost expects the initial model to be specified as a function with data as input.

  l = lrn("classif.rpart")

  init_predictor = function(data) {
    # Get response prediction from Learner
    p = l$predict_newdata(data)$response
    # One-hot encode and take first column

We can now run Multi-Calibration Boosting by instantiating the object and calling the multicalibrate method. Note, that typically, we would use Multi-Calibration on a separate validation set! We furthermore select the auditor model, a SubpopAuditorFitter, in our case a Decision Tree:

  mc = MCBoost$new(
    init_predictor = init_predictor,
    auditor_fitter = "TreeAuditorFitter")
  mc$multicalibrate(train_data, train_labels)

Lastly, we predict on new data.

tstid = setdiff(tsk$row_ids, tid) # held-out data
test_data = tsk$data(cols = tsk$feature_names, rows = tstid)


While mcboost in its defaults implements Multi-Accuracy (Kim et al., 2019), it can also multi-calibrate predictors (Hebert-Johnson et al., 2018). In order to achieve this, we have to set the following hyperparameters:

  mc = MCBoost$new(
    init_predictor = init_predictor,
    auditor_fitter = "TreeAuditorFitter",
    num_buckets = 10,
    multiplicative = FALSE

MCBoost as a PipeOp

mcboost can also be used within a mlr3pipeline in order to use at the full end-to-end pipeline (in the form of a GraphLearner).

  gr = ppl_mcboost(lrn("classif.rpart"))
  tsk = tsk("sonar")
  tid = sample(1:208, 108)
  gr$predict(tsk$clone()$filter(setdiff(1:208, tid)))

Further Examples

The mcboost vignettes Basics and Extensions and Health Survey Example demonstrate a lot of interesting showcases for applying mcboost.


This R package is licensed under the LGPL-3. If you encounter problems using this software (lack of documentation, misleading or wrong documentation, unexpected behaviour, bugs, …) or just want to suggest features, please open an issue in the issue tracker. Pull requests are welcome and will be included at the discretion of the maintainers.

As this project is developed with mlr3’s style guide in mind, the following resources can be helpful to individuals wishing to contribute: Please consult the wiki for a style guide, a roxygen guide and a pull request guide.

Code of Conduct

Please note that the mcboost project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

Citing mcboost

If you use mcboost, please cite our package as well as the two papers it is based on:

    author = {Pfisterer, Florian and Kern, Christoph and Dandl, Susanne and Sun, Matthew and 
    Kim, Michael P. and Bischl, Bernd},
    title = {mcboost: Multi-Calibration Boosting for R},
    journal = {Journal of Open Source Software},
    doi = {10.21105/joss.03453},
    url = {},
    year = {2021},
    publisher = {The Open Journal},
    volume = {6},
    number = {64},
    pages = {3453}
  # Multi-Calibration
    title = {Multicalibration: Calibration for the ({C}omputationally-Identifiable) Masses},
    author = {Hebert-Johnson, Ursula and Kim, Michael P. and Reingold, Omer and Rothblum, Guy},
    booktitle = {Proceedings of the 35th International Conference on Machine Learning},
    pages = {1939--1948},
    year = {2018},
    editor = {Jennifer Dy and Andreas Krause},
    volume = {80},
    series = {Proceedings of Machine Learning Research},
    address = {Stockholmsmässan, Stockholm Sweden},
    publisher = {PMLR}
  # Multi-Accuracy
    author = {Kim, Michael P. and Ghorbani, Amirata and Zou, James},
    title = {Multiaccuracy: Black-Box Post-Processing for Fairness in Classification},
    year = {2019},
    isbn = {9781450363242},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {},
    doi = {10.1145/3306618.3314287},
    booktitle = {Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society},
    pages = {247--254},
    location = {Honolulu, HI, USA},
    series = {AIES '19}