Sample locality preserving discriminant analysis for classification
Liping Yang, 辜小花 Gu Xiao-hua, 叶洪伟 YE Hong-wei · Optics and Precision Engineering · 2011
The small sample size and the loss of effective dimension problems always exist in discriminative dimension reduction methods of high-dimensional data classification.To address these problems,a Sample Locality Preserving Discriminant Analysis(SLPDA) method is proposed by integrating the latest patch alignment framework and Locality Preserving Projections(LPP).The within-class and out-class neighborhood relationships of all samples in the SLPDA are constructed by summing the within-class and out-class neighborhood graphs of each sample,respectively.Thereafter,the optimal mapping from a high-dimensional input space to a low-dimensional feature space of the SLPDA is obtained by making the within-class neighbors of all samples as close as possible and meanwhile keeping the out-class neighbors as distant as possible.The proposed SLPDA method avoids the small sample size problem of high-dimensional data classification and extends the effective dimension of low-dimensional feature space.Experimental results on several high-dimensional face databases,e.g.ORL,FERET and PIE,indicate that the proposed SLPDA method significantly outperforms the classical discriminative dimension reduction methods.Comparing with Discriminative Locality Alignment(DLA),which is also a dimension reduction method based on patch alignment framework,the recognition rate of SLPDA on a FERET subset is 4.5% higher.