Detect Noisy Label Based on Ensemble Learning

Abstract

The success of machine learning relies on high-quality labeled training data. If there are incorrectly labeled data in the training data, the performance of the best classifier will be greatly reduced in a wide range of classification problems, and noisy tags are also often more harmful than noisy attributes. Unfortunately, large datasets almost contain incorrect or inaccurate labels. This paper proposes a simple and effective method that can identify noisy data in text classification datasets to a certain extent. We analyze the characteristics of the noisy data, and design the new method based on the idea of ensemble learning. The method combines with the majority voting and iterative methods to select the noisy data hidden in the dataset. Under the same conditions, our method can select more noisy data, and perform corresponding evaluations on the recall and precision of noise. The experimental results show that this method is better than some previous methods.

This work was supported by the National Key R&D Program of China (Grant No: 2017YFB0803203) and Shanghai Municipal Natural Science Foundation (Grant No. 15ZR1403700).