Probabilistic classification: Difference between revisions

Revision as of 19:40, 10 September 2017

Latest revision as of 17:26, 19 August 2025

(31 intermediate revisions by 21 users not shown)

Line 1:

{{Short description|Machine learning problem}}

{{machine learning bar}}

In [[machine learning]], a '''probabilistic classifier''' is a [[statistical classification|classifier]] that is able to predict, given an observation of an input, a [[probability distribution]] over a [[Set (mathematics)|set]] of classes, rather than only outputting the most likely class that the observation should belong to. Probabilistic classifiers provide classification that can be useful in its own right<ref>{{cite book |first1=Trevor |last1=Hastie |first2=Robert |last2=Tibshirani |first3=Jerome |last3=Friedman |year=2009 |title=The Elements of Statistical Learning |url=http://statweb.stanford.edu/~tibs/ElemStatLearn/ |page=348 |quote=[I]n [[data mining]] applications the interest is often more in the class probabilities <math>p_\ell(x), \ell = 1, \dots, K</math> themselves, rather than in performing a class assignment. |url-status=dead |archive-url=https://web.archive.org/web/20150126123924/http://statweb.stanford.edu/~tibs/ElemStatLearn/ |archive-date=2015-01-26 }}</ref> or when combining classifiers into [[ensemble classifier|ensembles]].

==Types of classification==

Line 7 ⟶ 8:

:<math>\hat{y} = f(x)</math>

The samples come from some set {{mvar|X}} (e.g., the set of all [[document classification|documents]], or the set of all [[Computer vision#RecognitionBababoui|images]]), while the class labels form a finite set {{mvar|Y}} defined prior to training.

Probabilistic classifiers generalize this notion of classifiers: instead of functions, they are [[conditional probability|conditional]] distributions <math>\Pr(Y \vert X)</math>, meaning that for a given <math>x \in X</math>, they assign probabilities to all <math>y \in Y</math> (and these probabilities sum to one). "Hard" classification can then be done using the [[Bayes estimator|optimal decision rule]]<ref name="bishop">{{cite book |first=Christopher M. |last=Bishop |year=2006 |title=Pattern Recognition and Machine Learning |publisher=Springer}}</ref>{{rp|39–40}}

Line 15 ⟶ 16:

or, in English, the predicted class is that which has the highest probability.

Binary probabilistic classifiers are also called [[binomialbinary regression]] models in [[statistics]]. In [[econometrics]], probabilistic classification in general is called [[discrete choice]].

Some classification models, such as [[naive Bayes classifier|naive Bayes]], [[logistic regression]] and [[multilayer perceptron]]s (when trained under an appropriate [[loss function]]) are naturally probabilistic. Other models such as [[support vector machine]]s are not, but [[#Probability calibration|methods exist]] to turn them into probabilistic classifiers.

==Generative and conditional training==

Some models, such as [[logistic regression]], are conditionally trained: they optimize the conditional probability <math>\Pr(Y \vert X)</math> directly on a training set (see [[empirical risk minimization]]). Other classifiers, such as [[naive Bayes]], are trained [[Generative model|generatively]]: at training time, the class-conditional distribution <math>\Pr(X \vert Y)</math> and the class [[Prior probability|prior]] <math>\Pr(Y)</math> are found, and the conditional distribution <math>\Pr (Y \vert X)</math> is derived using [[Bayes' theorem|Bayes' rule]].<ref name="bishop"/>{{rp|43}}

==Probability calibration==

{{Main article|Calibration (statistics)}}

Not all classification models are naturally probabilistic, and some that are, notably naive Bayes classifiers, [[decision tree learning|decision trees]] and [[Boosting (machine learning)|boosting]] methods, produce distorted class probability distributions.<ref name="Niculescu">{{cite conference |last1=Niculescu-Mizil |first1=Alexandru |first2=Rich |last2=Caruana |title=Predicting good probabilities with supervised learning |conference=ICML |year=2005 |url=http://machinelearning.wustl.edu/mlpapers/paper_files/icml2005_Niculescu-MizilC05.pdf | doi = 10.1145/1102351.1102430 |url-status=dead |archive-url=https://web.archive.org/web/20140311005243/http://machinelearning.wustl.edu/mlpapers/paper_files/icml2005_Niculescu-MizilC05.pdf |archive-date=2014-03-11 }}</ref> In the case of decision trees, where {{math|Pr(''y''{{!}}'''x''')}} is the proportion of training samples with label {{mvar|y}} in the leaf where {{math|'''x'''}} ends up, these distortions come about because learning algorithms such as [[C4.5]] or [[Predictive analytics#Classification and regression trees|CART]] explicitly aim to produce homogeneous leaves (giving probabilities close to zero or one, and thus high [[Bias of an estimator|bias]]) while using few samples to estimate the relevant proportion (high [[Bias–variance tradeoff|variance]]).<ref>{{cite conference |first1=Bianca |last1=Zadrozny |first2=Charles |last2=Elkan |title=Obtaining calibrated probability estimates from decision trees and naive Bayesian classifiers |url=http://cseweb.ucsd.edu/~elkan/calibrated.pdf |year=2001 |conference=ICML |pages=609–616}}</ref>

[[File:Calibration plot.png|thumb|An example calibration plot]] Calibration can be assessed using a '''calibration plot''' (also called a '''reliability diagram''').<ref name="Niculescu" /><ref>{{Cite web|url=https://jmetzen.github.io/2015-04-14/calibration.html|title=Probability calibration|website=jmetzen.github.io|access-date=2019-06-18}}</ref> A calibration plot shows the proportion of items in each class for bands of predicted probability or score (such as a distorted probability distribution or the "signed distance to the hyperplane" in a support vector machine). Deviations from the identity function indicate a poorly-calibrated classifier for which the predicted probabilities or scores can not be used as probabilities. In this case one can use a method to turn these scores into properly [[Calibration (statistics)|calibrated]] class membership probabilities.

For the [[binary classification|binary]] case, a common approach is to apply [[Platt scaling]], which learns a [[logistic regression]] model on the scores.<ref name="platt99">{{cite journal |last=Platt |first=John |authorlinkauthor-link=John Platt (computer scientist) |title=Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods |journal=Advances in largeLarge marginMargin classifiersClassifiers |volume=10 |issue=3 |year=1999 |pages=61–74 |url=httphttps://www.researchgate.net/publication/2594015_Probabilistic_Outputs_for_Support_Vector_Machines_and_Comparisons_to_Regularized_Likelihood_Methods/file/504635154cff5262d6.pdf2594015}}</ref>

An alternative method using [[isotonic regression]]<ref>{{Cite book | last1 = Zadrozny | first1 = Bianca| last2 = Elkan | first2 = Charles| doi = 10.1145/775047.775151 | chapter = Transforming classifier scores into accurate multiclass probability estimates | chapter-url = httphttps://www.cs.cornell.edu/courses/cs678/2007sp/ZadroznyElkan.pdf| title = Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining - KDD '02 | pages = 694–699| year = 2002 | isbn = 978-1-58113-567-X| pmid = | pmc = 1| id = [[CiteSeerX]]: {{URL|1=citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.13.7457|2=10.1.1.13.7457}}| citeseerx = 10.1.1.164.8140| s2cid = 3349576}}</ref> is generally superior to Platt's method when sufficient training data is available.<ref name="Niculescu"/>

In the [[multiclass classification|multiclass]] case, one can use a reduction to binary tasks, followed by univariate calibration with an algorithm as described above and further application of the pairwise coupling algorithm by Hastie and Tibshirani.<ref>{{Cite journal | last1 = Hastie | first1 = Trevor| last2 = Tibshirani | first2 = Robert| doi = 10.1214/aos/1028144844 | title = Classification by pairwise coupling | journal = [[The Annals of Statistics]] | volume = 26 | issue = 2 | pages = 451–471| year = 1998 | pmid = | pmc = | zbl = 0932.62071| id = [[CiteSeerX]]: {{URL|1=citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.46.6032|2=10.1.1.46.6032}}| citeseerx = 10.1.1.309.4720 }}</ref>

==Evaluating probabilistic classification==

Commonly used lossevaluation functionsmetrics forthat compare the predicted probability to probabilisticobserved classificationoutcomes include [[log loss]] and the, [[meanBrier squared errorscore]] between the predicted, and thea truevariety probabilityof calibration distributionserrors. The former of these is commonlyalso used toas traina loss function in the training of logistic models.

Calibration errors metrics aim to quantify the extent to which a probabilistic classifier's outputs are ''well-calibrated''. As [[Philip Dawid]] put it, "a forecaster is well-calibrated if, for example, of those events to which he assigns a probability 30 percent, the long-run proportion that actually occurs turns out to be 30 percent".<ref>{{cite journal |doi=10.1080/01621459.1982.10477856 |title=The Well-Calibrated Bayesian |journal=Journal of the American Statistical Association |volume=77 |issue=379 |pages=605–610 |year=1982 |last1=Dawid |first1=A. P}}</ref> Foundational work in the ___domain of measuring calibration error is the Expected Calibration Error (ECE) metric.<ref>{{cite book | first1 = M.P. | last1= Naeini | first2 = G. | last2 = Cooper| first3 = M. | last3 = Hauskrecht| chapter = Obtaining well calibrated probabilities using bayesian binning |title = Proceedings of the AAAI Conference on Artificial Intelligence | year = 2015 | chapter-url = https://www.dbmi.pitt.edu/wp-content/uploads/2022/10/Obtaining-well-calibrated-probabilities-using-Bayesian-binning.pdf}}</ref> More recent works propose variants to ECE that address limitations of the ECE metric that may arise when classifier scores concentrate on narrow subset of the [0,1], including the Adaptive Calibration Error (ACE) <ref>{{cite book | first1 = J. | last1= Nixon | first2 = M.W. | last2 = Dusenberry| first3 = L. | last3 = Zhang| first4= G. | last4 = Jerfel | first5 = D. | last5 = Tran | chapter = Measuring Calibration in Deep Learning |title = CVPR workshops | year = 2019 | chapter-url = https://openaccess.thecvf.com/content_CVPRW_2019/papers/Uncertainty%20and%20Robustness%20in%20Deep%20Visual%20Learning/Nixon_Measuring_Calibration_in_Deep_Learning_CVPRW_2019_paper.pdf}}</ref> and Test-based Calibration Error (TCE).<ref>{{cite book | first1 = T. | last1= Matsubara | first2 = N. | last2 = Tax| first3 = R. | last3 = Mudd| first4= I. | last4 = Guy | chapter = TCE: A Test-Based Approach to Measuring Calibration Error |title = Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence (UAI) | year = 2023 | arxiv= 2306.14343 }}</ref>

A method used to assign scores to pairs of predicted probabilities and actual discrete outcomes, so that different predictive methods can be compared, is called a [[scoring rule]].

==Software Implementations==

* MoRPE<ref>{{cite web |title=MoRPE |url=https://github.com/adaviding/morpe |website=GitHub |access-date=17 February 2023}}</ref> is a trainable probabilistic classifier that uses [[isotonic regression]] for probability calibration. It solves the [[multiclass classification|multiclass]] case by reduction to binary tasks. It is a type of kernel machine that uses an inhomogeneous polynomial kernel.

==References==