基于Bayes的有噪训练集去噪方法研究

Identifying and Correcting Mislabled Training Instances Using Bayes

利用有噪训练集训练分类器的过程中,去噪是基本的预处理步骤。传统的去噪工作只是简单地删除被标记为噪声的实例。显然,这样处理会清除噪声实例中的有用信息。本文提出一种基于Bayes的去噪方法,不但能辨识出噪声而且能纠正噪声实例的错误类标,从而保证其有效信息不会丢失。

De-noising is a basic pretreatment in the process of training a classifier. Most traditional de-noising approaches only delete instances tagged as noise which obviously also eliminates the useful information in these instances. A new approach is presented with which we can not only identify noise but also correct it, so that the useful information will be reserved.