摘要
核独立分量分析(KICA)可以实现特征提取,但当数据量较大时,特征的数量也随之增加.针对这种情况,提出了一种结合类可分性和遗传算法来选择特征并降低其维数的方法.对数据进行KICA处理后得到特征向量及权值矩阵.对于权值矩阵,使用类间类内距离比来进行特征初选.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.对这些特征向量使用遗传算法来选择最优特征组.两个实验验证了该方法的有效性.
KICA can implement feature extraction,but the number of the features rises along with the data.For the case,a methods combining the class separability and genetic algorithm is proposed for feature election and dimension reduction.Firstly,the feature vectors and weight vectors of original data are obtained by KICA.Then,the distance ratio of between-class and within-class is used for feature selection.The columns of the weight matrix with larger distance ratio and the corresponding features 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.
出处
《武汉理工大学学报（交通科学与工程版）》
2009年第4期772-775,共4页
Journal of Wuhan University of Technology（Transportation Science & Engineering）
基金
国家民委自然科学基金项目资助(批准号:m209012)
关键词
核独立分量分析(KICA)
特征选择
类可分性
距离度量
遗传算法
kernel independent component analysis（KICA）
feature selection
class separability
distance measurement
genetic algorithm
