Abstract：For the classification learning tasks, it is inevitable to generate noise in the process of acquiring data. Especially, the presence of label noise may not only make the learning model more complex, but also lead to overfitting and poor generalization ability. Although some proposed label noise filtering algorithms can solve the above problems to some extent, there are still some limitations such as poor noise recognition ability, unsatisfactory classification effect and low filtering efficiency et al. Focused on the issue, a local probability sampling method based on label confidence distribution for label noise filtering was proposed. Firstly, it applies the random forest classifiers to vote on the labels of the samples and obtains the confidence of label of each sample. And then all samples will be divided into easy identification and difficult identification ones according to the level of labels confidence. Finally, they will be processed with different filtering strategies, respectively. The experimental results show that in the presence of label noise, the method proposed can maintain high noise recognition ability in most cases, and the classification generalization performance also has obvious advantages.
