稀疏表示保持的鉴别特征选择算法

摘要

稀疏表示作为一种基于部分数据的表示,已经吸引了越来越多的关注,并广泛应用于模式识别和机器学习领域.提出一种新的算法,称为稀疏表示保持的鉴别特征选择(SRPFS),其目的是选择鉴别性特征子集,使得在所选特征子空间中,样本的稀疏类内重构残差和稀疏类间重构残差的差值最小化.与传统算法选择特征的独立性方式不同,该算法以批处理方式选择最具鉴别性的特征,并用于优化提出的l2,1范数最小化的目标函数.在标准UCI数据集和哥伦比亚图像数据库的实验结果表明,该算法在识别性能和稳定性方面优于其他经典特征选择算法.%A new algorithm for selection of distinguishable features preserved by sparse representation,whose aim is to se-lect a subset of distinguishable features to minimize the difference value of reconstruction residual in sparse class and reconstruc-tion residuals between sparse classes of samples in the subspace of selected features. The algorithm,which is different from the selection feature independence mode of the traditional algorithms,selects the most distinguishable features in batch mode and, is used to optimize the minimized objective function of l2,1-norm. The experimental results on standard UCI datasets and Colum-bia object image data base show that the algorithm is superior to other classic feature selection algorithms in the aspects of recog-nition performance and stability.