一种基于类内类间距离的ICA特征选择方法

Feature Selection Method Based on ICA and Distance Ratio of within/between Class

独立分量分析(ICA)可以实现特征提取,但不能直接用于特征选择。对数据进行ICA后得到混合矩阵和独立分量,独立分量可以作为特征矢量,混合矩阵可以用于进行特征选择。首先,使用一种距离度量来计算混合矩阵每一类的类内类间距离比;然后对每一类按该比值由小到大重新排列混合矩阵和独立分量,保留权重矩阵中类间类内距离比大的列,及其对应的特征向量;最后对这些特征向量使用遗传算法选择最优特征组。两个实验验证了该方法的有效性。

ICA can implement feature extraction, but it can＇t directly fit for feature selection. The results after ICA are mixing matrix and independent components, the latter is regarded as feature vectors, while the mixing matrix can be used for feature selection. First, the distance ratio between within - class and between - class of each column of mixing matrix is computed by using a distance measurement. Then, for each class, the mixing matrix and independent components are realigned by sort as cending according to the distance rate,and the columns of the weight matrix with smaller distance ratio and the corresponding featur.es are reserved. Finally,the best feature set is selected by genetic algorithm from these foregoing feature vectors. Two experiments show that the proposed method is valid.