文本分类中基于散度差的线性特征抽取方法

Method of linear feature extraction based on scatter difference in text categorization

分析了特征选择与特征抽取的特点与不足,针对Fisher线性鉴别准则存在问题,在一种加权散度差线性鉴别准则的基础上提出了一种基于散度差与SVD相结合的文本特征抽取方法。在解决了类内散布矩阵的奇异性问题困扰同时,通过对低阶矩阵的奇异值分解取代了对高阶矩阵的特征值求解,计算量大大减少。在最低限度减少信息损失的前提下实现了特征维数的大幅度减缩。试验结果表明,这种方法在文本分类上的准确性较好。

The advantage and disadvantage about feature selection and extraction in text categorization is analyzed. In allusion to the problems of Fisher criterion, a weighted scatter difference method is brought forward. And more, a new method that combined the scatter difference and SVD is presented to realize feature extraction. Meanwhile, solving the problem that within-class scatter matrix Sw is singular, the SVD is realized in few dimension matrix and need not to calculate the eigenvalue in more dimension matrix. So the difficulty of calculation is lower. At the precondition of lower information loss we reduce the feature dimension. Lastly, we have a test about text categorization and the result shows that this method has a better precision.