基于依赖最大化和稀疏回归的多标签特征选择  

Multi-label feature selection with dependence maximization and sparse regression

作　　者：吴喆君 黄睿[1] WU Zhe-jun;HUANG Rui(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China)

机构地区：[1]上海大学通信与信息工程学院,上海200444

出　　处：《计算机工程与设计》2022年第7期1898-1904,共7页Computer Engineering and Design

摘　　要：针对基于稀疏回归的多标签特征选择方法中数据的特征和标签之间线性关系假设不成立的问题,提出一种基于依赖最大化和稀疏回归的多标签特征选择方法(multi-label feature selection with dependence maximization and sparse regression,DMSR)。构建数据的低维子空间,最大化低维空间与数据的标签空间之间的依赖性,使用希尔伯特-施密特独立性准则作为依赖性的计算依据,将数据从特征空间映射到该低维空间,设计一种交替优化的算法对稀疏回归模型进行求解,得到用于特征选择的投影矩阵。在多个不同类型的多标签数据集上的实验结果表明,所提算法的性能优于其它对比算法。Aiming at the problem that the assumption of the linear relationship between features and labels of data does not hold in multi-label feature selection methods based on the sparse regression,a multi-label feature selection method with dependence maximization and sparse regression(DMSR)was proposed.A low-dimensional subspace of the data was constructed and the dependence between the low-dimensional space and the label space was maximized.Hilbert-Schmidt independence criterion was adopted to measure the dependence.The data were mapped from the feature space to the low-dimensional space.An alternating optimization algorithm was designed to solve the sparse regression model and to obtain the mapping matrix for feature selection.Experimental results on various multi-label data sets demonstrate that the proposed method outperforms other comparing methods.

关 键 词：多标签学习 特征选择 依赖最大化 稀疏回归 低维空间 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]