The impact of diversity on the accuracy of evidential classifier ensembles

Research output: Contribution to journalArticle

45 Citations (Scopus)

Abstract

Diversity being inherent in classifiers is widely acknowledged as an important issue in constructingsuccessful classifier ensembles. Although many statistics have been employed inmeasuring diversity among classifiers to ascertainwhether it correlates with ensemble performancein the literature, most of these measures are incorporated and explained in anon-evidential context. In this paper, we provide a modeling for formulating classifier outputsas triplet mass functions and a uniform notation for defining diversity measures. Wethen assess the relationship between diversity obtained by four pairwise and non-pairwisediversitymeasures and the improvement in accuracy of classifiers combined in different ordersby Demspter’s rule of combination, Smets’ conjunctive rule, the Proportion and Yager’srules in the framework of belief functions. Our experimental results demonstrate that theaccuracy of classifiers combined by Dempster’s rule is not strongly correlated with thediversity obtained by the four measures, and the correlation between the diversity andthe ensemble accuracy made by Proportion and Yager’s rules is negative, which is not infavor of the claim that increasing diversity could lead to reduction of generalization error ofclassifier ensembles.
LanguageEnglish
Pages584-607
JournalInternational Journal of Approximate Reasoning
Volume53
Issue number4
DOIs
Publication statusPublished - Jun 2012

Fingerprint

modeling
statistics

Cite this

@article{bac6b59edbee417fb13913ffe876a77b,
title = "The impact of diversity on the accuracy of evidential classifier ensembles",
abstract = "Diversity being inherent in classifiers is widely acknowledged as an important issue in constructingsuccessful classifier ensembles. Although many statistics have been employed inmeasuring diversity among classifiers to ascertainwhether it correlates with ensemble performancein the literature, most of these measures are incorporated and explained in anon-evidential context. In this paper, we provide a modeling for formulating classifier outputsas triplet mass functions and a uniform notation for defining diversity measures. Wethen assess the relationship between diversity obtained by four pairwise and non-pairwisediversitymeasures and the improvement in accuracy of classifiers combined in different ordersby Demspter’s rule of combination, Smets’ conjunctive rule, the Proportion and Yager’srules in the framework of belief functions. Our experimental results demonstrate that theaccuracy of classifiers combined by Dempster’s rule is not strongly correlated with thediversity obtained by the four measures, and the correlation between the diversity andthe ensemble accuracy made by Proportion and Yager’s rules is negative, which is not infavor of the claim that increasing diversity could lead to reduction of generalization error ofclassifier ensembles.",
author = "Yaxin Bi",
year = "2012",
month = "6",
doi = "10.1016/j.ijar.2011.12.011",
language = "English",
volume = "53",
pages = "584--607",
journal = "International Journal of Approximate Reasoning",
issn = "0888-613X",
publisher = "Elsevier",
number = "4",

}

The impact of diversity on the accuracy of evidential classifier ensembles. / Bi, Yaxin.

In: International Journal of Approximate Reasoning, Vol. 53, No. 4, 06.2012, p. 584-607.

Research output: Contribution to journalArticle

TY - JOUR

T1 - The impact of diversity on the accuracy of evidential classifier ensembles

AU - Bi, Yaxin

PY - 2012/6

Y1 - 2012/6

N2 - Diversity being inherent in classifiers is widely acknowledged as an important issue in constructingsuccessful classifier ensembles. Although many statistics have been employed inmeasuring diversity among classifiers to ascertainwhether it correlates with ensemble performancein the literature, most of these measures are incorporated and explained in anon-evidential context. In this paper, we provide a modeling for formulating classifier outputsas triplet mass functions and a uniform notation for defining diversity measures. Wethen assess the relationship between diversity obtained by four pairwise and non-pairwisediversitymeasures and the improvement in accuracy of classifiers combined in different ordersby Demspter’s rule of combination, Smets’ conjunctive rule, the Proportion and Yager’srules in the framework of belief functions. Our experimental results demonstrate that theaccuracy of classifiers combined by Dempster’s rule is not strongly correlated with thediversity obtained by the four measures, and the correlation between the diversity andthe ensemble accuracy made by Proportion and Yager’s rules is negative, which is not infavor of the claim that increasing diversity could lead to reduction of generalization error ofclassifier ensembles.

AB - Diversity being inherent in classifiers is widely acknowledged as an important issue in constructingsuccessful classifier ensembles. Although many statistics have been employed inmeasuring diversity among classifiers to ascertainwhether it correlates with ensemble performancein the literature, most of these measures are incorporated and explained in anon-evidential context. In this paper, we provide a modeling for formulating classifier outputsas triplet mass functions and a uniform notation for defining diversity measures. Wethen assess the relationship between diversity obtained by four pairwise and non-pairwisediversitymeasures and the improvement in accuracy of classifiers combined in different ordersby Demspter’s rule of combination, Smets’ conjunctive rule, the Proportion and Yager’srules in the framework of belief functions. Our experimental results demonstrate that theaccuracy of classifiers combined by Dempster’s rule is not strongly correlated with thediversity obtained by the four measures, and the correlation between the diversity andthe ensemble accuracy made by Proportion and Yager’s rules is negative, which is not infavor of the claim that increasing diversity could lead to reduction of generalization error ofclassifier ensembles.

U2 - 10.1016/j.ijar.2011.12.011

DO - 10.1016/j.ijar.2011.12.011

M3 - Article

VL - 53

SP - 584

EP - 607

JO - International Journal of Approximate Reasoning

T2 - International Journal of Approximate Reasoning

JF - International Journal of Approximate Reasoning

SN - 0888-613X

IS - 4

ER -