查询结果:   赵迎利,王凯明,肖玉柱,宋学力.基于l惩罚典型相关分析的特征选择[J].计算机应用与软件,2019,36(10):279 - 284.
中文标题
基于l惩罚典型相关分析的特征选择
发表栏目
算法
摘要点击数
262
英文标题
FEATURE SELECTION BASED ON l1,2 PENALIZED CANONICAL CORRELATION ANALYSIS
作 者
赵迎利 王凯明 肖玉柱 宋学力 Zhao Yingli Wang Kaiming Xiao Yuzhu Song Xueli
作者单位
长安大学理学院 陕西 西安 710064     
英文单位
School of Science, Changan University, Xian 710064, Shaanxi, China     
关键词
特征选择 典型相关分析 l1,2 范数 随机分组
Keywords
Feature selection Canonical correlation analysis l1,2 -norm Random grouping
基金项目
长安大学中央高校基本科研业务费专项资金项目(310812163504)
作者资料
赵迎利,硕士生,主研领域:机器学习。 王凯明,副教授。肖玉柱,副教授。宋学力,教授。 。
文章摘要
特征选择是多模态高维数据机器学习的一个热点问题,而过拟合和过稀疏是特征选择需要克服的关键问题。对此提出以l1,2 范数作为惩罚项兼顾稀疏作用和光滑作用,以组内稀疏来防止过拟合,以组间光滑来防止过稀疏,通过优化数据间的相关性来实现特征选择。然而对一般数据而言,群组信息又很难获得,所以对于群组信息缺失的数据,应用随机分组获得群组信息,最终实现兼顾组间光滑和组内稀疏优点的特征选择。模拟实验结果表明,该方法能较完整地选择出两模态数据间的关联特征,并且去除不相关特征。
Abstract
In machine learning of multi-modal and high-dimensional data, feature selection is a hot research issue, and over-fitting and over-sparseness are the key problem needed to be solved. Focusing on this problem, we proposed l1,2 norm as penalty term to take into account both sparsity and smoothness, to prevent over-fitting by intra-group sparsity, to prevent over-sparsity by inter-group smothness, and to achieve feature selection by optimizing the correlation between data. However, for general data, the group information is difficult to obtain, so for the data without given group information, the random-grouping method was used to obtain it, and finally feature-selection was realized considering both inter-group smoothness and intra-group sparsity. The simulation results show that the proposed method can select features completely and remove the unrelated features as well.
下载PDF全文