Research On Parameters Adaptively Setting Manifold Learning Algorithms Based On Sparse Representation

Manifold learning is one of the most popular issues of machine learning.It is also a representative of nonlinear dimensionality reduction.Manifold learning algorithms perform effectively for nonlinear data sets dimensionality reduction and feature extraction.But manifold learning algorithms are also of some limitations.For example,most manifold learning algorithms need tedious process to adjust parameter for neighborhood selection.Because most manifold learning algorithms reduce the dimensionality of the original data by locality preserving,which causes the problem of how to assign the parameter of neighborhood size.In most cases,it will be set manually.In order to solve the problem,this thesis proposes two parameters adaptively setting manifold learning algorithms based on sparse representation.The main contributions of this thesis are listed below:(1)A parameter adaptively setting discriminant projection method based on group sparse is proposed.In this method,group sparse is used to describe the geometry structure of data sets.In this way,the neighborhood graph can be constructed adaptively,which naturally avoid the weakness of setting parameter by manual operation in traditional manifold learning methods.In addition,both a local sparse scatter and a non-local sparse scatter are also proposed to unify sparse representation and locality theory into a mathematical model and improve the discriminant ability of this algorithm.(2)The supervised adaptively parameter setting discriminant projection method is also proposed.Compared to the original algorithm,this approach proposes a supervised sparse representation framework which can describe the geometry structure of data sets more accurately by using the label information of data sets to construct sparse matrix.It can also limit the interference of pseudo-neighbor points.In addition,the intra-class scatter and inter-class scatter are introduced into this method to promote discriminative ability.The results of the experiments on some common face databases indicate that parameter adaptively setting discriminant projection method based on group sparse have higher recognition rate compared to some state-of-art methods.In addition,the supervised parameter adaptively setting discriminant projection method have better recognition rate than the first method.