基于双稀疏表示的无监督属性选择算法  

An Unsupervised Feature Selection Algorithm Based on Double Sparse Representation

作　　者：劳翠金[1] 秦燊[1] 文国秋 LAO Cui-jin;QIN Shen;WEN Guo-qiu(Information Engineering Department,Liuzhou City Vocational College,Liuzhou 545000,China;School of Computer Science and Information Technology,Guangxi Normal University,Guilin 541004,China)

机构地区：[1]柳州城市职业学院信息工程系,广西柳州545000 [2]广西师范大学计算机科学与信息工程学院,广西桂林541004

出　　处：《控制工程》2021年第4期774-780,共7页Control Engineering of China

基　　金：国家自然科学基金资助项目(61573270)。

摘　　要：由于大量无类标签数据需要降维处理,近年来无监督属性选择学习受到越来越多的关注。通过将两种稀疏表示和属性自表达损失函数嵌入到同一个学习框架中,提出了一种新的无监督属性选择算法。首先,利用属性自表达技术重构数据,使每一重构属性为所有原始属性的线性表征,加强属性关联性。然后,分别利用l_(2, p)范数正则项和l_(1)范数正则项使权重系数矩阵稀疏,剔除冗余无关属性,实现属性选择目的。最后,将约简后的低维数据集送入支持向量机中获得分类结果,以此评判属性选择算法是否有效。对多个真实数据集进行实验,实验结果显示,所提算法的降维效果优于一般常用算法。Since a large amount of unlabeled data need to reduce dimension, unsupervised feature selection learning has been attracting more and more attention in recent years. This paper proposes a novel unsupervised feature selection algorithm via embedding two sparse representations and feature-level self-representation loss function into a unified learning framework. Firstly, the algorithm utilizes feature-level self-representation to reconstruct original data to make each reconstructed feature be linearly represented by all original features, enhancing their correlation. Then, l_(2, p)-norm and l_(1)-norm regularization terms are used simultaneously to make the reconstruction coefficient matrix sparse, aiming at unselecting redundant and irrelevant features, and achieving feature selection. Finally, the dimension-reduced data sets are fed into support vector machine(SVM) to yield classification accuracy that is the evaluation metric for performance of feature selection algorithms. The experimental results on real data sets indicate that our proposed algorithm outperforms the commonly used dimension-reduced algorithms.

关 键 词：属性选择 稀疏表示 重构技术 属性自表达 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]