CNC-NOS: Class noise cleaning by ensemble filtering and noise scoring
作者：
Highlights：
• A novel class noise cleaner able not only to remove noisy instances but to successfully relabel them.
• It introduces an improvement of the noise score measure already proposed in a recent filter, making it sensitive to the proximity of the neighbors.
• A wide range of noise levels are used to compare the proposal in order to validate its behavior in different scenarios.
• Experimental results outperform the state-of-the-art approaches in noise filtering and other noise correctors for class noise, the most common approach.
• A thorough analysis on the effect of the compared approaches in terms of successfully and wrongly treated instances is provided and related to its performance.
摘要
•A novel class noise cleaner able not only to remove noisy instances but to successfully relabel them.•It introduces an improvement of the noise score measure already proposed in a recent filter, making it sensitive to the proximity of the neighbors.•A wide range of noise levels are used to compare the proposal in order to validate its behavior in different scenarios.•Experimental results outperform the state-of-the-art approaches in noise filtering and other noise correctors for class noise, the most common approach.•A thorough analysis on the effect of the compared approaches in terms of successfully and wrongly treated instances is provided and related to its performance.
论文关键词：Classification,Data preprocessing,Noisy data,Class noise,Noise filtering,Data reparation,Data relabeling
论文评审过程：Received 1 May 2017, Revised 17 October 2017, Accepted 21 October 2017, Available online 27 October 2017, Version of Record 6 December 2017.
论文官网地址：https://doi.org/10.1016/j.knosys.2017.10.026
