Study Of Classification Models For Noisy Data

Research has shown that the widely used boosting algorithm Ada Boost suffers from over-fitting when dealing with noisy data sets due to noise sensitivity limitations.Recently,a representative approach named noise detection based Adaboost(ND-Ada Boost)has been proposed to solve the noise sensitivity problem of the Ada Boost.The idea of ND-Ada Boost algorithm is to make the classifier notice noisy samples when classifying them and work on misclassifying mislabelled samples as a way to improve the generalisation ability of the classifier.However,the algorithm simply and brutally detects the samples as normal and noisy,so that the classifier cannot be biased in its classification.That is a classifier cannot do a better job of selecting two mislabeled samples when misclassified at the same time.Based on this idea,this paper proposes a new sample noise detection method and adds the sample noise probability values to the loss function,derives new sample update weights and new base classifier coefficients,and proposes a MKBoost algorithm named noise of sample attributes-detection based MKBoost(NSAD-MKB)that with stronger generalization capability for data containing noise.The algorithm MKBoost is an algorithm proposed based on the Ada Boost algorithm framework,which can effectively solve nonlinear classification problems.Moreover,the training error of the proposed loss function is calculated in this paper,and it is concluded that the loss function of NSAD-MKB algorithm decreases at an exponential rate during the training process.Finally,data from the UCI was selected to illustrate through experimental results that NSAD-MKB outperforms ND-Ada Boost on the test dataset.