Research On Stochastic Coordinate Algorithm Of Support Vector Machines And Robust Support Vector Machines Under The Background Of Big Data

In the current era of big data,SVM is confronted with various new challenges.First of all,the large-scale training data brings huge challenges to the storage of SVM kernel matrix,so it is necessary to design a more efficient algorithm to solve large-scale SVM problems.Secondly,the high frequency updating of data forces SVM to have better stability for training samples,however,the selection of parameter in the classical SVM model is more sensitive to training samples,so it is necessary to study algorithm for solving more stable SVM based on Ivanov regularization term(I-SVM).Thirdly,the noise and outliers of large-scale real data bring new challenges to the robustness of SVM.More and more researchers also pay attention to the research of robust SVM.In addition,the diversity of big data structure requires SVM to be able to handle the data with distinct structures,however,the current research on the algorithm of multiple kernel learning(MKL),which is suitable for dealing with diverse data,needs to be developed and improved urgently.Finally,the low value density of big data requires SVM to select features rapidly and cost-effectively.SCAD-SVM(smooth clipped absolute deviation)has good Oracle statistical properties in feature selection,but because of its nonconvexity,it's significant to have a further study on nonconvex regularization term SVM.With the above challenges of SVM in the era of big data,this thesis conducts research on the following aspects:(1)The algorithm of large-scale SVM and the linear convergence rate.In order to solve the problem of large-scale kernel matrix storage,we design a random primal-dual coordinate algorithm to solve large-scale SVM.The iterative process of the algorithm has a simple closed form and occupies less computer memory.Furthermore,the linear convergence condition of the algorithm is given in the thesis.The numerical results show that our algorithm is more efficient than the libsvm solver which is widely used at present.(2)The algorithm framework of I-SVM.The algorithm framework of I-SVM is designed to overcome the shortcomings of current I-SVM algorithm.In this thesis,a general framework for solving I-SVM is given,which is suitable for large-scale sample calculation.The numerical results show that the algorithm in this thesis is more effective than the existing algorithm.(3)Global robust SVM and distributed robust SVM algorithm.In order to overcome the noise and outliers in the real data,based on the idea of global robust optimization and distributed robust optimization,two kinds of new robust SVM are proposed,and their respective equivalent computable forms(counterparts)are given.The numerical results show that the two kinds of robust SVM proposed in this thesis have better generalization ability than the traditional SVM.(4)Large scale non-sparse multiple kernel learning algorithm.As a single kernel learning classifier,traditional SVM has a weaker ability to deal with diverse data problems than the multiple kernel learning(MKL)classifier.Under the background of big data,a random coordinate descent algorithm is designed to solve non-sparse multiple kernel learning.The iterative process of the algorithm has a closed form.We combine fuzzy clustering with multiple kernel learning and propose a local multiple kernel learning classifier.The experimental results show that the local multiple kernel learning classifier proposed by us is better than the existing main classifier.(5)The algorithm for high dimensional SCAD-SVM.Mining valuable information from big data with low value density requires SVM to have the ability of feature selection.Therefore,SCAD-SVM model is studied and an effective algorithm is designed to solve SCADSVM.The algorithm iterative process in this thesis has a closed form.The experimental results show that the algorithm proposed in this thesis is effective.In the background of big data,this thesis studies the effective algorithm of large-scale SVM.The research results enrich the SVM method and theory,and provide more effective learning and decision-making tools for practical application.