Improved label noise identification by exploiting unlabeled data

Abstract

In machine learning, the available training samples are not always perfect and some labels can be corrupted which are called label noises. This may cause the reduction of accuracy. Meanwhile it will also increase the complexity of model. To mitigate the detrimental effect of label noises, noise filtering has been widely used which tries to identify label noises and remove them prior to learning. Almost all existing works only focus on the mislabeled training dataset and ignore the existence of unlabeled data. In fact, unlabeled data are easily accessible in many applications. In this work, we explore how to utilize these unlabeled data to increase the noise filtering effect. To this end, we have proposed a method named MFUDCM (Multiple Filtering with the aid of Unlabeled Data using Confidence Measurement). This method applies the novel multiple soft majority voting idea to make use unlabeled data. In addition, MFUDCM is expected to have a higher accuracy of identifying mislabeled data by using the concept of multiple voting. Finally, the validity of the proposed method MFUDCM is confirmed by experiments and the comparison results with other methods.