Classification Of High Dimensional Data Based On Dimensionality Reduction And Kernel Learning

Recently,with the rapid development of information technology,image,audio and video,text and bioinformatics data tend to show high dimensional feature of the Internet,the high dimension of the input data will often lead to the existing machine learning model performance decline or ill posed problems,how to effectively extract features from high-dimensional data,improve the performance of the task specific learning in high dimensional data environment,has become a hot topics of pattern recognition,machine learning,data mining,computer visionand so on.In view of the existing high-dimensional data classification problem in high dimension,nonlinear,small sample and other challenges,the theory and method of learning,pattern recognition and computer interdisciplinary machine utilization,with dimensionality reduction,multiple kernel learning,nonparametric kernel learning,sparse representation based method,aiming at the shortcomings of the present stage of high dimensional data dimensionality reduction algorithm and kernel learning algorithm,high dimensional data recognition method of kernel learning and based on the sparse representation,and the proposed method has been successfully applied in face recognition,text classification and other practical problems.The studies mainly include the following four achievements:(1)To improve the problem of the existing multi kernel dimensionality reduction algorithms learning efficiency is not high,we introduced the spectral regression method into the traditional multi kernel dimensionality reduction model and formula derivation,and proposed multi-kernel dimensionality reduction algorithms based on spectral regression and trace ratio maximum.We also proposed more efficient and effective solution to the generalized characteristic of dense matrix multi kernel dimensionality reduction model in value decomposition based on the design of the proposed model.The experimental results on high dimensional data sets show that the proposed method has the advantages of both spectral regression and multi-kernel learning and outperforms state-of-the-art multi-kernel dimensionality reduction algorithms.(2)The existing supervised multi-kernel dimensionality reduction need to meet the mandatory conditionsy that is each class of data are subject to strong constraints Gauss distribution.In order to solve the problem when the condition is not established for high dimensional data,this thesis proposed spectral regression based marginal fisher analysis algorithm.We proposed linear,sigle and muti-kernel dimensionality reduction algorithm based on MFA.The training algorithm has the advantages of spectral regression and MFA,which can make full use of the manifold structure of the data set and the category information,and can solve the problem of high dimension reduction under the condition of non Gauss distribution of the original data.(3)When dealing with high dimensional and sparse data with nonparametric kernel learning in semi-supervised,the manifold assumption may failure.In order to solve this problem,we proposed a semi-supervised low rank non-parametric kernel learning method and proved that the model can be converted to trace ratio optimization problems.To overcome the dependence on manifold assumption,we deduced a series of semi supervised kernel learning algorithm embedded low rank based on the proposed framework.Experimental results on benchmark datasets demonstrate that the proposed method outperforms other state-of-the art algorithms.(4)In view of the low classification accuracy and high computational complexity of traditional kernel sparse representation classification algorithms on high dimensional data sets,a kernel sparse representation classification method based on spectral regression is proposed.Sparse representation is an important branch in the field of machine learning,and is widely used in face recognition.Firstly,spectral regression analysis is used to obtain the transformation matrix for feature extraction,and the sample data are extracted by the transformation matrix.Then the kernel method will project the features to high dimensional feature space to make it more separability.Finally the use of kernel method in high dimensional feature space enables us to use the sparse representation method to identify the face images.By combining the spectral regression method with the kernel sparse representation classification method,the manifold structure and class information of the data set are effectively utilized,which can solve the problem of sparse representation classification of high dimension face images.Experimental results on the standard face image data sets show that the proposed method not only improves the recognition rate,but also reduces the time cost of the algorithm.It can be applied in the classification of high dimensional face image data effectively.