Beyond sigmoids: How to obtain well-calibrated probabilities from binary classifiers with beta calibration
Open accessThe DOI matches its Crossref record; 14 of the 16 references it cites were found in a registry or an index, and 2 are not checked yet.
Abstract
For optimal decision making under variable class distributions and misclassification costs a classifier needs to produce well-calibrated estimates of the posterior probability. Isotonic calibration is a powerful non-parametric method that is however prone to overfitting on smaller datasets; hence a parametric method based on the logistic sigmoidal curve is commonly used. While logistic calibration is designed for normally distributed per-class scores, we demonstrate experimentally that many classifiers including Naive Bayes and Adaboost suffer from a particular distortion where these score distributions are heavily skewed. In such cases logistic calibration can easily yield probability estimates that are worse than the original scores. Moreover, the logistic curve family does not include the identity function, and hence logistic calibration can easily uncalibrate a perfectly calibrated classifier. In this paper we solve all these problems with a richer class of parametric calibration maps based on the beta distribution. We derive the method from first principles and show that fitting it is as easy as fitting a logistic curve. Extensive experiments show that beta calibration is superior to logistic calibration for a wide range of classifiers: Naive Bayes, Adaboost, random forest, logistic regression, support vector machine and multi-layer perceptron. If the original classifier is already calibrated, then beta calibration learns a function close to the identity. On this we build a statistical test to recognise if the model deviates from being well-calibrated.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Beta calibration is equivalent to logistic calibration if 𝛽 1 = 𝛽 2 .8,9 Otherwise, it is more flexible than logistic calibration.”
From Calibrating machine learning approaches for probability estimation: A comprehensive comparison (2023) · cites this paper as [8]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“One approach that works well is to model the classwise scores with Beta distributions (Kull et al., 2017a(Kull et al., , 2017b)).”
From Classifier calibration: a survey on how to assess and improve predicted class probabilities (2023) · cites this paper as Kull et al., 2017b
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Isotonic calibration 8,12 is one of the most commonly used calibration approaches by the machine learning community.64”
From Calibrating machine learning approaches for probability estimation: A comprehensive comparison (2023) · cites this paper as [8]
The cited source says what the claim attributes to it.
In the citing paper
“One approach that works well is to model the classwise scores with Beta distributions (Kull et al., 2017a(Kull et al., , 2017b)).”
From Classifier calibration: a survey on how to assess and improve predicted class probabilities (2023) · cites this paper as Kull et al., 2017a
The cited source says what the claim attributes to it.
In the citing paper
“Beta calibration is specifically designed for probabilistic classifiers and alleviates both drawbacks of Platt scaling (Kull et al., 2017a, b).”
From Classifier calibration: a survey on how to assess and improve predicted class probabilities (2023) · cites this paper as Kull et al., 2017a, b
2 more claims about this paper
Create a free account to see every claim other papers make about this one.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Scikit-learn: Machine Learning in Python
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Müller, A., Nothman, J., Louppe, G., et al. (2012) · arXiv · doi:10.48550/arxiv.1201.0490 · Cited by 71,016 · cited 4×
“The second is the one implemented in Python’s Scikit-learn, based on Adaboost method SAMME (Zhu et al., 2009), referred to as Ada-S.”
MethodInfluentialThe WEKA data mining software: an update
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., & Witten, I. H. (2009) · ACM SIGKDD Explorations Newsletter · doi:10.1145/1656274.1656278 · Cited by 17,750 · cited 1×
“However, as the task is simply univariate logistic regression with feature s and label y , it can be solved using the standard logistic regression functionality in any machine learning toolkit, including WEKA (Hall et al., 2009) and Scikit-learn (Pedregosa et al., 2011).”
MethodFriedman, J., Hastie, T., & Tibshirani, R. (2000) · The Annals of Statistics · doi:10.1214/aos/1016218223 · Cited by 6,884
Predicting good probabilities with supervised learning
Niculescu-Mizil, A., & Caruana, R. (2005) · Proceedings of the 22nd international conference on Machine learning - ICML '05 · doi:10.1145/1102351.1102430 · Cited by 1,884 · cited 2×
“This was expected, since isotonic calibration is the only non-parametric method in this study and literature has shown that it needs more data than logistic calibration to fit a good calibration map (Niculescu-Mizil and Caruana, 2005).”
MethodHastie, T., Rosset, S., Zhu, J., & Zou, H. (2009) · Statistics and Its Interface · doi:10.4310/sii.2009.v2.n3.a8 · Cited by 1,875 · cited 1×
“The second is the one implemented in Python’s Scikit-learn, based on Adaboost method SAMME (Zhu et al., 2009), referred to as Ada-S.”
MethodTransforming classifier scores into accurate multiclass probability estimates
Zadrozny, B., & Elkan, C. (2002) · Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining · doi:10.1145/775047.775151 · Cited by 1,280 · cited 3×
“The scores of a sufficiently calibrated classifier can be simply thresholded at a threshold directly derived from the misclassification cost (Zadrozny and Elkan, 2002).”
BackgroundMethodRobust classification for imprecise environments
Provost, F., & Fawcett, T. (2001) · Machine Learning · doi:10.1023/a:1007601015854 · Cited by 1,250
Platt, J. C. (2000) · Advances in Large-Margin Classifiers · doi:10.7551/mitpress/1113.003.0008 · Cited by 1,156 · cited 3×
“Logistic calibration was proposed by (Platt, 2000) to reduce overfitting by introducing a strong inductive bias which considers only calibration maps of the following form: where γ, δ are real-valued parameters with γ ≥ 0 to ensure that the calibration map is monotonically non-decreasing.”
MethodStatistical Comparisons of Classifiers over Multiple Data Sets
DemšarJanez (2006) · Journal of Machine Learning Research · doi:10.5555/1248547.1248548 · Cited by 1,074
Calibration of Probabilities: The State of the Art
Lichtenstein, S., Fischhoff, B., & Phillips, L. D. (1977) · Decision Making and Change in Human Affairs · doi:10.1007/978-94-010-1276-8_19 · Cited by 483
Kull, M., Filho, T. M. S., & Flach, P. A. (2017) · Bristol Research (University of Bristol) · Cited by 371
Self-citationProperties and Benefits of Calibrated Classifiers
Cohen, I., & Goldszmidt, M. (2004) · Lecture notes in computer science · doi:10.1007/978-3-540-30116-5_14 · Cited by 125 · cited 2×
“Formally, a scoring classifier is perfectly calibrated on a dataset if for each of its output scores s the proportion of positives within instances with model output score s is equal to s (Cohen and Goldszmidt, 2004).”
BackgroundFawcett, T., & Niculescu-Mizil, A. (2007) · Machine Learning · doi:10.1007/s10994-007-5011-0 · Cited by 109 · cited 2×
“Well-established calibration methods in binary classification include logistic calibration, also known as ‘Platt scaling’ in reference to the author who introduced it for support vector machines (Platt, 2000) and isotonic calibration, also known as the ROC convex hull method and pool-adjacent-violators (Zadrozny and E…”
MethodForecast aggregation via recalibration
Turner, B. M., Mark Steyvers, B., Merkle, E. C., Budescu, D. V., & Wallsten, T. S. (2013) · Machine Learning · doi:10.1007/s10994-013-5401-4 · Cited by 87 · cited 1×
“The two-parameter version of beta calibration with a and c where a = b has been considered before as a linear-in-log-odds (LLO) calibration method (Lichtenstein, Fischhoff and Phillips, 1977; Turner et al., 2014) but without a justification.”
MethodImproving the AUC of Probabilistic Estimation Trees
Ferri, C., Flach, P. A., & Hernández‐Orallo, J. (2003) · Lecture notes in computer science · doi:10.1007/978-3-540-39857-8_13 · Cited by 77
Self-citationKull, M., & Flach, P. (2015) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-23528-8_5 · Cited by 43 · cited 1×
“…beta[m= 1 / 2 ] calibration, isotonic calibration, logistic calibration and uncalibrated probabilities, in terms of Brier score (BS) and log-loss (LL), which are both proper scoring rules (Kull and Flach, 2015) and hence well-founded measures for assessing the quality of predicted probabilities.”
BackgroundSelf-citation
Cited by
Classifier calibration: a survey on how to assess and improve predicted class probabilities
Silva Filho, T., Song, H., Perello-Nieto, M., Santos-Rodriguez, R., Kull, M., & Flach, P. (2023) · Machine Learning · doi:10.1007/s10994-023-06336-7 · Cited by 228
“( ) Similar cross-validation training is also adopted as proposed in Kull et al. (2017a).”
BackgroundMethodInfluentialSelf-citationKull, M., Perello-Nieto, M., Kängsepp, M., Filho, T. S., Song, H., & Flach, P. (2019) · arXiv · doi:10.48550/arxiv.1910.12656 · Cited by 206
“For 2 classes, the Dirichlet calibration map family coincides with the beta calibration map family [15].”
MethodSelf-citationVerified Uncertainty Calibration
Kumar, A., Liang, P., & Ma, T. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1909.10155 · Cited by 131
“We first describe the scaling family we use, which is Platt scaling after applying a log-transform [12], otherwise known as beta calibration [45].”
MethodPerformance Evaluation in Machine Learning: The Good, the Bad, the Ugly, and the Way Forward
Flach, P. A. (2019) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v33i01.33019808 · Cited by 221
Self-citationIdentifying and prioritizing potential human-infecting viruses from their genome sequences
Mollentze, N., Babayan, S. A., & Streicker, D. G. (2021) · PLoS Biology · doi:10.1371/journal.pbio.3001390 · Cited by 138
Metadata
- Authors
- Kull, M., Silva Filho, T. M., & Flach, P.
- Affiliations
- University of Tartu (EE)
- University of Bristol (GB)
- Venue
- Electronic Journal of Statistics
- Publisher
- Institute of Mathematical Statistics
- Published
- 2017-01-01
- Volume
- 11
- Issue
- 2
- Pages
- 5052-5080
- ISSN
- 1935-7524 (ISSN-L)
- OpenAlex
- W2773235741
- Cited by
- 111
- References
- 15
- Citations by year
- 2026: 22 · 2025: 16 · 2024: 11 · 2023: 16 · 2022: 10 · 2021: 17 · 2020: 11 · 2019: 7 · 2018: 1
- FWCI
- 3.10
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- Locations
- Electronic Journal of Statistics (publishedVersion, cc-by)
- Explore Bristol Research (submittedVersion, cc-by)
- Project Euclid (Cornell University) (submittedVersion)
- Bristol Research (University of Bristol) (submittedVersion, cc-by)
- Funders
- Eesti Teadusagentuur (PUT1458)
- Conselho Nacional de Desenvolvimento Científico e Tecnológico
- Engineering and Physical Sciences Research Council (EP/K031910/1, EP/K018728/1)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Imbalanced Data Classification Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Keywords
- probability calibration, binary classifiers, posterior probability estimation, isotonic calibration, naive Bayes, AdaBoost, random forest, support vector machine, multilayer perceptron, calibration tests
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.