A Novel Noise Filter Based on Multiple Voting

Abstract

Label noises exist in many applications, which usually add difficulties for data analysis. A straightforward and effective method is to detect and filter out them prior to training. Ensemble learning based filter has shown promising performances. We define an important parameter to improve the performance of the algorithm. The proposed method is cost sensitive which integrates the mislabeled training dataset and noise costs for learning. Finally, the experimental results on the benchmark datasets show the superiority of the proposed method.