期刊名称：International Journal of Computer Science and Network Security
印刷版ISSN：1738-7906
出版年度：2018
卷号：18
期号：8
页码：148-152
出版社：International Journal of Computer Science and Network Security
摘要：Real-world datasets are not perfect and always suffer from noise that may affect classifiers built under the effect of such type of disturbance. Different types of noise are existing in almost any real-world problem, but not always known. Existence of noise decreases the accuracy of a classifier and increases its training time and complexity of the induced model. Most of existing machine learning algorithms have integrated different approaches to enhance their learning abilities in presence of noise, but noise still can make negative impacts. Therefore noise robustness of a classifier is an important issue in noisy environments and should be studied. This paper evaluates the robustness of different machine learning algorithms against class noise. The Equalized Loss of Accuracy (ELA) is the robustness metric which is used in this study. Ten benchmark datasets with 0-20% of noise level are used in experiments and finally ELA results of algorithms are compared.
关键词：;;;; ;;;;;; Noise impact; Classification; Robustness metric; Class noise
