关键词: 标签噪声, 动态概率抽样, 噪声过滤, 标签置信度, 置信度

Abstract: In machine learning problems, data quality will have a profound impact on the accuracy of the prediction of system. Because of the difficulty in obtaining information and the subjective and limited cognition of human beings, experts cannot mark all samples accurately. In recent years, some probabilistic sampling methods can't avoid the unreasonable and subjective problem of artificial division of samples. To solve this problem, this paper puts forward a label noise filtering method based on dynamic probability sampling, which fully considers the differences among samples of each data set, analyzes the trend of information entropy of confidence distribution in each interval, and determines a reasonable threshold. Experimental results show that the method proposed in this paper can obtain higher label noise recognition ability and classification ability in most cases.

Key words: label noise, dynamic probability sampling, noise filtering, label confidence, confidence