基于谱技巧稀疏保持判别分析(英文)

SparsityPreservingDiscriminant Analysis Based on Spectral Techniques

稀疏保持判别分析(SPDA)是一种新型的基于图的半监督降维(SSDR)方法,近年来已被成功应用于解决诸多实际问题(如人脸识别).SPDA基于数据的稀疏重构关系建图,由于稀疏的特性,从而包含自然的判别信息.然而,在SPDA计算中涉及到稠密矩阵的特征分解从而导致在存储和记忆方面会耗费大量时间,为此,我们提出了一种新的SSDR算法-基于谱技巧稀疏保持判别分析(SSPDA),该方法将稀疏表示与谱技巧结合在一起.具体地,首先把投影函数的计算转化为一个回归类优化问题,然后借助岭回归技巧得到投影方向向量,从而有效的避免了稠密矩阵的特征分解问题.在两个单标号人脸数据上的实验结果表明了该算法的有效性.

parsity Preserving Discriminant Analysis（SPDA） is a popular graph-based semi-supervised dimensionality reduction（SSDR） algorithm,which does not only contain natural discriminating information by sparse reconstruction relationship of the data sets,but also more applicable to face recognition problem with only a few training sample.However,the computation of SPDA involves eigen-decomposition of the dense matrix which is expensive in both memory and computation.In this paper,we propose an improved SPDA algorithm called Sparsity Preserving Discriminant Analysis based on Spectral Techniques（SSPDA）.Specifically,the proposed approach casts discriminant analysis into a regression framework,and then the projection directions can be efficiently computed by related ridge regression,thus eigen-decomposition of the dense matrix is avoid.Experimental results on single training image face recognition demonstrate the effectiveness of the proposed algorithm.