Research On Typical Correlation Analysis Algorithm Of Multiple Sets Based On Sparse Representation

Canonical correlation analysis (CCA) is an effective algorithm for feature extraction, which is used in correlation problem between two groups of feature data in the pattern recognition. In the real world, the same pattern always exhibits various of characteristics. Therefore, this article embarks from the fusion of multi group feature. The multiset canonical correlation analysis(MCCA) algorithm based on sparse representation is the foundation of the research. Then, we combine with the kernel technology, class information, maximizing margin criterion and other related basic theory knowledge. We set up a series of algorithm, feature extraction and feature fusion. In this paper, the specific research and innovation work includes the following parts:(1)Due to the canonical correlation analysis algorithm based on sparse preserving projection can not meet the needs of the nonlinear problem in the practical work, and inspired by kernel technology, we research the canonical correlation analysis algorithm based on kernel sparse preserving projection(KSPCCA). In order to avoid the optimal selection of kernel function in the choose of kernel and fuse multi group feature we research a simple multiset canonical correlation analysis algorithm based on the multiple kernel sparse preserving projection according to the multiple kernel learning idea and multiset canonical correlation analysis algorithm based on sparse preserving projection.(2) We solve the regularization equation of the sparse representation coefficients with L2 norm. We can get the collaborative representation coefficients. We establish the sample reconstructing adjacency with the collaborative representation coefficients to extract subspace feature. Because the idea of the collaborative representation projection doesn’t consider the class information, then, we add class information into the collaborative representation projection, we propose the theory of the collaborative representation discriminative (CRD). Then combining with MCCA and the kernel method, we research the multiset canonical correlation analysis algorithm based on collaborative representation projection (CRMCCA), the multiset canonical correlation analysis algorithm based on kernel collaborative representation projection(KCRMCCA), the multiset canonical correlation analysis algorithm based on collaborative representation discriminative (CDMCCA) and the multiset canonical correlation analysis algorithm based on kernel collaborative representation discriminative (KCDMCCA).(3) In solving the sparse representation coefficients, we only consider the similar samples, then we can obtain sparse coefficient between similar samples. At the same time, combining with the maximizing margin criterion (MMC), we propose the discriminant sparse neighborhood preserving embedding algorithm(DSNPE). With the idea, we propose the multiset canonical correlation analysis algorithm based on discriminant sparse neighborhood preserving embedding algorithm (DSNPEMCCA).