An Adaptive Voting Mechanism Based on Relative Density for Filtering Label Noises
Longhai Huang, Yabin Shao, Jialin Peng · 2022 IEEE 5th International Conference on Electronics Technology (ICET) · 2022
Label noise is an important problem in classification. As an efficient method to deal with label noise, the filtering method does not need to estimate the noise rate and does not depend on any loss function. In the relative density (RD) model, the relative density value can be used as a feature to detect noise, but the efficiency of filtering noise in the algorithm depends entirely on the setting of the threshold, which makes the algorithm not generalizable. In order to solve the threshold problem in the RD model, an adaptive voting mechanism based on relative density values is proposed in this paper. In this model, instead of setting a fixed threshold value for the model, the suspected noise is screened out by cross-validation, and then the noise is detected and filtered by the voting mechanism. And through a series of experiments, the effectiveness of our method is demonstrated.