基于相对贡献率的噪声裁剪算法

Noise Clipping Algorithm Based on Relative Contribution Rate

提出了一种基于相对贡献率的噪声裁剪算法(Class noise cutting, CNC)。通过计算得到特征对于主题的相对贡献率,利用特征区分度评分挑选对于当前主题分类最有价值的特征集,选出相应的候选类别,减少候选类别集,提高了分类准确率,加快了分类器的响应速度。与另一种噪声裁剪算法(Eliminating class noise, ECN)比较,CNC具有更高的准确率,由于具有更精简的特征维度词典以及更优异的候选类别集使得响应速度大大加快。

This paper presents a class noise cutting algorithm(Class noise cutting, CNC) based on relative contribution rate. The algorithm calculates the relative contribution rate of features to the theme. The most valuable feature set is selected by using features distinguish rating. The corresponding candidate categories for each feature are selected, to reduece the candidate category set, improves the classification accuracy, and speed up the response speed of the classifier. Compared with another ECN noise cutting algorithm(Eliminating the class whose), CNC-has higher accuracy and because of its simpler feature dimension dictionary and better candidate category set, the response speed is greatly accelerated.