Class noise detection using classification filtering algorithms

Abstract

One of the significant problems in classification is class noise which has numerous potential consequences such as reducing the overall accuracy and increasing the complexity of the induced model. Subsequently, finding and eliminating misclassified instances are known as important phases in machine learning and data mining. The predictions of classifiers can be applied to detect noisy instances, inconsistent data and errors, what is called classification filtering. It creates a new set of dataset to develop a reliable and precise classification model. In this paper we analyze the effect of class noise on six supervised learning algorithms. To evaluate the performance of the classification filtering algorithms, several experiments were conducted on six real datasets. Finally, the noisy instances are removed and relabeled and the performance was then measured using evaluation criteria. The findings of this study show that classification filtering have a potential capability to detect class noise.