Title : ( Ensemble of online neural networks for non-stationary and imbalanced data streams )
Authors: Adel Ghazikhani , Reza Monsefi , Hadi Sadoghi Yazdi ,
دانلود فایل برای اعضای دانشگاهAccess to full-text not allowed by authors
Abstract
Concept drift(non-stationarity)andclassimbalancearetwoimportantchallengesforsupervised classifiers. “Concept drift” (or non-stationarity)referstochangesintheunderlyingfunctionbeinglearnt, and classimbalanceisavastdifferencebetweenthenumbersofinstancesindifferentclassesofdata. Class imbalanceisanobstaclefortheefficiency ofmostclassifiers. Researchonclassification ofnon- stationary andimbalanceddatastreams,mainlyfocusesonbatchsolutions,whereasonlinemethodsare more appropriate.Here,weproposeanonlineensembleofneuralnetwork(NN)classifiers. Ensemble models arethemostfrequentmethodsusedforclassifyingnon-stationaryandimbalanceddatastreams. The maincontributionisatwo-layerapproachforhandlingclassimbalanceandnon-stationarity.Inthe first layer,cost-sensitivelearningisembeddedintothetrainingphaseoftheNNs,andinthesecondlayer a newmethodforweightingclassifiers oftheensembleisproposed.Theproposedmethodisevaluated on 3syntheticand8real-worlddatasets.Theresultsshowstatisticallysignificant improvement compared toonlineensemblemethodswithsimilarfeatures.
Keywords
, Data streamclassification Online ensemble Concept drift Imbalanced data Cost, sensitivelearning
@article{paperid:1037393,
author = {Ghazikhani, Adel and Monsefi, Reza and Sadoghi Yazdi, Hadi},
title = {Ensemble of online neural networks for non-stationary and imbalanced data streams},
journal = {Neurocomputing},
year = {2013},
volume = {122},
number = {12},
month = {December},
issn = {0925-2312},
pages = {535--544},
numpages = {9},
keywords = {Data streamclassification
Online ensemble
Concept drift
Imbalanced data
Cost-sensitivelearning},
}
%0 Journal Article
%T Ensemble of online neural networks for non-stationary and imbalanced data streams
%A Ghazikhani, Adel
%A Monsefi, Reza
%A Sadoghi Yazdi, Hadi
%J Neurocomputing
%@ 0925-2312
%D 2013
