基于双向稀疏的多视图子空间学习算法  

A MULTI-VIEW SUBSPACE LEARNING METHOD BASED ON BIDIRECTIONAL SPARSITY

作　　者：杨凡[1] 饶雨泰[1] Yang Fan;Rao Yutai(School of Software Engineering,Hubei Radio and TV University,Wuhan 430074,Hubei,China)

机构地区：[1]湖北广播电视大学软件工程学院,湖北武汉430074

出　　处：《计算机应用与软件》2023年第6期266-275,共10页Computer Applications and Software

基　　金：湖北省教育厅科学技术研究项目(B2016592)。

摘　　要：针对传统的多视图子空间学习方法很难找到一个有效的子空间维数并同时处理异常值,提出一种基于双向稀疏的多视图子空间学习方法。将通过矩阵分解得到的低维表示分解成两个矩阵,在第一个矩阵中加入l_(p,2)行稀疏范数来捕获相关数据之间没有冗余特征的良好共享特征,即数据的二次特征提取;为了同时识别离群点并减少其影响,在第二个矩阵中加入l_(p,2)列稀疏矩阵。进一步提出一种有效的求解算法,并分析了该算法的收敛性以及计算复杂度。在多个数据集上仿真结果表明,与传统的多视图无监督特征提取方法相比,该方法能够更加有效地解决子空间维数问题,并且对异常值情形具有较强的鲁棒性。Aimed at the problem that the traditional multi-view subspace learning methods are difficult to find an effective subspace dimension and deal with outliers at the same time,a multi-view subspace learning method based on bidirectional sparsity is proposed.The low dimensional representation obtained by matrix decomposition was decomposed into two matrices.l p,2 row sparse norm was added to the first matrix to capture the good shared features without redundant features between related data,that was,the secondary feature extraction of data.In order to identify outliers at the same time and reduce their influence,l p,2 column sparse matrix was added to the second matrix.Furthermore,an effective algorithm was proposed and its convergence and computational complexity were analyzed.Simulation results on multiple datasets show that the proposed method can solve the subspace dimension problem more effectively than the traditional multi-view unsupervised feature extraction methods,and has strong robustness to outliers.

关 键 词：多视图 子空间学习 双向稀疏性 鲁棒性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]