Robust Streaming Data Classification Based On Noise Filtration And Correction

Classification of noisy streaming data is an important topic in streaming data mining,and it has attracted a lot of attention due to its reality and generality.Existing classification methods suffer from some drawbacks when dealing with noise: some methods do not handle the noise at all which makes the prediction accuracy of classifier influenced by the noise;other methods handle the noise but fail to consider attribute noise and label noise at the same time,leading to low identification rate of noise and limited feasibility.To solve this problem,this thesis proposes a new noisy streaming data classification method NFC(Noise Filtration and Correction),aiming at improving the prediction accuracy of classifier given the presence of both attribute noise and label noise.NFC models the attribute probability and label probability of samples,and then estimates the noise samples via maximizing the attribute probability and label probability,and filters attribute noise and corrects label noise at last.In addition,the training of weights of classifiers is integrated in the estimation of noise,and thus the ensemble of classifiers and the estimation of noise are combined into one,which leads to more robust prediction results.By considering attribute noise and label noise separately,NFC estimates the two kinds of noise independently and handle them respectively,to insure that the processing of one kind of noise will not be influenced by another,which helps to improve the identification rate of each kind of noise.Empirical evaluations demonstrate that comparing to existing noisy streaming data classification methods NFC can identify the noise more accurately with only a little time cost,and retain more beneficial information,which improves the prediction accuracy of classifiers.