Neighborhood Classifier for Label Noise

One typical case of label noise indicates that some samples have been incorrectly labeled in data. Label noise of training samples will significantly affect the learning performances such that the classification accuracy will be reduced. Presently, many results of identifying samples of incorrect labels have been proposed. Most of them are based on the consideration of classifier based accuracy. Therefore, the performance of used classifier is directly related to the result of filtering samples with noise label. In this paper, a neighborhood strategy is introduced into analyzing label noise data, it is mainly because such classifier is superior to several popular classifiers. Not only the neighborhood classifier based algorithm is designed to remove samples with noise label, but also such type of filter is compared with the nearest neighborhood based filter. The experimental results demonstrate that our neighborhood classifier based filter performs well because higher classification accuracy can be achieved. This study suggests new trends for considering neighborhood approach to complex data.