基于核稀疏表示的特征选择算法  

Feature selection based on kernel sparse representation

作　　者：邓战涛[1] 胡谷雨[1] 潘志松[1] 张艳艳[1] 

机构地区：[1]解放军理工大学指挥自动化学院,南京210007

出　　处：《计算机应用研究》2012年第4期1282-1284,共3页Application Research of Computers

基　　金：国家自然科学基金资助项目(60603029);国家"863"计划资助项目(2010AAJ163)

摘　　要：为了解决高维数据在分类时导致的维数灾难,降维是数据预处理阶段的主要步骤。基于稀疏学习进行特征选择是目前的研究热点。针对现实中大量非线性可分问题,借助核技巧,将非线性可分的数据样本映射到核空间,以解决特征的非线性相似问题。进一步对核空间的数据样本进行稀疏重构,得到原数据在核空间的一种简洁的稀疏表达方式,然后构建相应的评分机制选择最优子集。受益于稀疏学习的自然判别能力,该算法能够选择出保持原始数据结构特性的"好"特征,从而降低学习模型的计算复杂度并提升分类精度。在标准UCI数据集上的实验结果表明,其性能上与同类算法相比平均可提高约5%。In order to solve the problem of dimension illness in classification of high-dimensional data,dimensionality reduction is a key approach in pretreatment.Feature selection based on sparse representation is one of the hottest research topics recently.In the face of the nonlinear problem,this paper motivated by kernel trick,nonlinear data was mapped into kernel space in which the nonlinear similarity of the features could be captured and reconstructed by sparse representation to get concision expression of original data in kernel space.Then it designed evaluate mechanism to select excellent feature subsets.As the natural discriminative power of sparse representation,＂good＂ feature which preserved the original structure would be selected so that feature selection could reduce computational complexity and improve precision.The results of experiment in standard UCI data sets show that the performance compare with similar algorithms improves about the average of 5%.

关 键 词：特征选择 稀疏表示 核技巧 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]