On PAC-Bayesian bounds for random forests

Stephan S. Lorenzen; Christian Igel; Yevgeny Seldin

doi:10.1007/s10994-019-05803-4

On PAC-Bayesian bounds for random forests

Stephan S. Lorenzen, Christian Igel^*, Yevgeny Seldin

^*Corresponding author af dette arbejde

Abstract

Existing guarantees in terms of rigorous upper bounds on the generalization error for the original random forest algorithm, one of the most frequently used machine learning methods, are unsatisfying. We discuss and evaluate various PAC-Bayesian approaches to derive such bounds. The bounds do not require additional hold-out data, because the out-of-bag samples from the bagging in the training process can be exploited. A random forest predicts by taking a majority vote of an ensemble of decision trees. The first approach is to bound the error of the vote by twice the error of the corresponding Gibbs classifier (classifying with a single member of the ensemble selected at random). However, this approach does not take into account the effect of averaging out of errors of individual classifiers when taking the majority vote. This effect provides a significant boost in performance when the errors are independent or negatively correlated, but when the correlations are strong the advantage from taking the majority vote is small. The second approach based on PAC-Bayesian C-bounds takes dependencies between ensemble members into account, but it requires estimating correlations between the errors of the individual classifiers. When the correlations are high or the estimation is poor, the bounds degrade. In our experiments, we compute generalization bounds for random forests on various benchmark data sets. Because the individual decision trees already perform well, their predictions are highly correlated and the C-bounds do not lead to satisfactory results. For the same reason, the bounds based on the analysis of Gibbs classifiers are typically superior and often reasonably tight. Bounds based on a validation set coming at the cost of a smaller training set gave better performance guarantees, but worse performance in most experiments.

Originalsprog	Engelsk
Tidsskrift	Machine Learning
Vol/bind	108
Udgave nummer	8-9
Sider (fra-til)	1503-1522
ISSN	0885-6125
DOI	https://doi.org/10.1007/s10994-019-05803-4
Status	Udgivet - 15 sep. 2019

Adgang til dokumentet

10.1007/s10994-019-05803-4

s10994-019-05803-4.pdfForlagets udgivne version, 644 KB

Andre filer og links

Link to publication in Scopus

Citationsformater

@article{e3ddb68a9c3d4961b72afaee07a187b6,

title = "On PAC-Bayesian bounds for random forests",

abstract = "Existing guarantees in terms of rigorous upper bounds on the generalization error for the original random forest algorithm, one of the most frequently used machine learning methods, are unsatisfying. We discuss and evaluate various PAC-Bayesian approaches to derive such bounds. The bounds do not require additional hold-out data, because the out-of-bag samples from the bagging in the training process can be exploited. A random forest predicts by taking a majority vote of an ensemble of decision trees. The first approach is to bound the error of the vote by twice the error of the corresponding Gibbs classifier (classifying with a single member of the ensemble selected at random). However, this approach does not take into account the effect of averaging out of errors of individual classifiers when taking the majority vote. This effect provides a significant boost in performance when the errors are independent or negatively correlated, but when the correlations are strong the advantage from taking the majority vote is small. The second approach based on PAC-Bayesian C-bounds takes dependencies between ensemble members into account, but it requires estimating correlations between the errors of the individual classifiers. When the correlations are high or the estimation is poor, the bounds degrade. In our experiments, we compute generalization bounds for random forests on various benchmark data sets. Because the individual decision trees already perform well, their predictions are highly correlated and the C-bounds do not lead to satisfactory results. For the same reason, the bounds based on the analysis of Gibbs classifiers are typically superior and often reasonably tight. Bounds based on a validation set coming at the cost of a smaller training set gave better performance guarantees, but worse performance in most experiments.",

keywords = "Majority vote, PAC-Bayesian analysis, Random forests",

author = "Lorenzen, {Stephan S.} and Christian Igel and Yevgeny Seldin",

year = "2019",

month = sep,

day = "15",

doi = "10.1007/s10994-019-05803-4",

language = "English",

volume = "108",

pages = "1503--1522",

journal = "Machine Learning",

issn = "0885-6125",

publisher = "Springer",

number = "8-9",

}

TY - JOUR

T1 - On PAC-Bayesian bounds for random forests

AU - Lorenzen, Stephan S.

AU - Igel, Christian

AU - Seldin, Yevgeny

PY - 2019/9/15

Y1 - 2019/9/15

N2 - Existing guarantees in terms of rigorous upper bounds on the generalization error for the original random forest algorithm, one of the most frequently used machine learning methods, are unsatisfying. We discuss and evaluate various PAC-Bayesian approaches to derive such bounds. The bounds do not require additional hold-out data, because the out-of-bag samples from the bagging in the training process can be exploited. A random forest predicts by taking a majority vote of an ensemble of decision trees. The first approach is to bound the error of the vote by twice the error of the corresponding Gibbs classifier (classifying with a single member of the ensemble selected at random). However, this approach does not take into account the effect of averaging out of errors of individual classifiers when taking the majority vote. This effect provides a significant boost in performance when the errors are independent or negatively correlated, but when the correlations are strong the advantage from taking the majority vote is small. The second approach based on PAC-Bayesian C-bounds takes dependencies between ensemble members into account, but it requires estimating correlations between the errors of the individual classifiers. When the correlations are high or the estimation is poor, the bounds degrade. In our experiments, we compute generalization bounds for random forests on various benchmark data sets. Because the individual decision trees already perform well, their predictions are highly correlated and the C-bounds do not lead to satisfactory results. For the same reason, the bounds based on the analysis of Gibbs classifiers are typically superior and often reasonably tight. Bounds based on a validation set coming at the cost of a smaller training set gave better performance guarantees, but worse performance in most experiments.

AB - Existing guarantees in terms of rigorous upper bounds on the generalization error for the original random forest algorithm, one of the most frequently used machine learning methods, are unsatisfying. We discuss and evaluate various PAC-Bayesian approaches to derive such bounds. The bounds do not require additional hold-out data, because the out-of-bag samples from the bagging in the training process can be exploited. A random forest predicts by taking a majority vote of an ensemble of decision trees. The first approach is to bound the error of the vote by twice the error of the corresponding Gibbs classifier (classifying with a single member of the ensemble selected at random). However, this approach does not take into account the effect of averaging out of errors of individual classifiers when taking the majority vote. This effect provides a significant boost in performance when the errors are independent or negatively correlated, but when the correlations are strong the advantage from taking the majority vote is small. The second approach based on PAC-Bayesian C-bounds takes dependencies between ensemble members into account, but it requires estimating correlations between the errors of the individual classifiers. When the correlations are high or the estimation is poor, the bounds degrade. In our experiments, we compute generalization bounds for random forests on various benchmark data sets. Because the individual decision trees already perform well, their predictions are highly correlated and the C-bounds do not lead to satisfactory results. For the same reason, the bounds based on the analysis of Gibbs classifiers are typically superior and often reasonably tight. Bounds based on a validation set coming at the cost of a smaller training set gave better performance guarantees, but worse performance in most experiments.

KW - Majority vote

KW - PAC-Bayesian analysis

KW - Random forests

UR - http://www.scopus.com/inward/record.url?scp=85065773116&partnerID=8YFLogxK

U2 - 10.1007/s10994-019-05803-4

DO - 10.1007/s10994-019-05803-4

M3 - Journal article

AN - SCOPUS:85065773116

SN - 0885-6125

VL - 108

SP - 1503

EP - 1522

JO - Machine Learning

JF - Machine Learning

IS - 8-9

ER -

On PAC-Bayesian bounds for random forests

Abstract

Adgang til dokumentet

Andre filer og links

Fingeraftryk

Citationsformater