面向不平衡文本的特征选择方法

在分析了传统特征选择方法构造的4项基本信息元素的基础上提出一种强类别信息的度量标准,并在此基础上,提出一种适用于不平衡文本的特征选择方法。该方法综合考虑了类别信息因子、词频因子,分别用于提高少数类和多数类类别分类精度。该方法在reuter-21578数据集上进行了实验,实验结果表明,该特征选择方法比IG、CHI方法都更好,不但微平均指标有一定程度的提高,而且宏平均指标也有一定程度的提高。

Feature Selection Method on Imbalanced Text

After analyzing the four basic information elements of traditional feature selection methods, a new measurement of strong class information is introduced and a new feature selection method is proposed for imbalanced text classification. The strong class information and the frequency of terms are used to improve the classification performance of minority classes and majority classes respectively. The experiments on reuter-21578 dataset show that the proposed method is better than IG and CHI. Both Micro F1 and Macro F1 are improved to some degree.