Font Size: a A A

Research On Support Vector Machine Models And Algorithms Based On Structural Information

Posted on:2018-08-02Degree:DoctorType:Dissertation
Country:ChinaCandidate:Q L HouFull Text:PDF
GTID:1319330515484186Subject:Strategy and management
Abstract/Summary:PDF Full Text Request
Support vector machine(SVM),which is recognized as an effective tool for classification and regression,has been widely used in many fields.As an extension of SVM,nonparallel hyperplane support vector machine(NHSVM),which receives extensive attention of many researchers,has become a new research hot spot duing to its better generalization ability and less computation complexity.In this paper,we excavate the potential structural information and discriminative information about the training samples to build more effective SVM and NHSVM models,and apply them to solve the practical problems.The main contents are summarized as follows:1.Inspired by linear discriminant analysis,this paper presents a novel nonparallel classifier termed as projection nonparallel support vector machine(PNPSVM).The new classifier,which is fully different from the existing nonparallel classifiers,needs two steps to obtain the two optimal proximal hyperplanes.The first step is to obtain two projection directions which could achieve maximum class separability by minimizing the within-class distance and maximizing the between-class distance.The second step is to determine the specific location of the two corresponding optimal proximal hyperplanes by choosing an appropriate central sample based on the underlying spatial distribution.In addition,the efficient successive overrelaxation algorithm is used to solve the optimization problems to speed up the solving procedure.The experimental results on both artificial dataset and UCI benchmark datasets verify the rationality and validity of our proposed approach.2.The conventional twin support vector machine(TWSVM)does not take into account the potential spatial distribution information and discriminative information about the training samples.In order to overcome this drawback,this paper proposes a novel classifier termed as twin support vector machine with improved pairwise constraints(IPCTSVM).Concretely,for each proximal hyperplane,we set up an corresponding optimization problem by introducing two novel discriminative regularization terms,i.e.,the intra-class discriminative regularization term and the inter-class discriminative regularization term,which consider the tightness between the similar patterns and the discrepancy between the dissimilar pairs,respectively.The new classifier is expected to not only learn the prior discriminative information about each constrained pair,but also combine the discrimination metric from the traditional pairwise constraints and the spatial distance measure from the heat kernel together.In the numerical experiment,the comprehensive experimental results on both artificial dataset and UCI benchmark datasets demonstrate the rationality and validity of our proposed approach.3.The standard SVM does not take into account the potential high-order relationships between different samples and discriminative information about each constrained pair.In this paper,by introducing a newly-designed discriminative regularization term,we propose a novel classifier termed as support vector machine with hypergraph-based pairwise constraints(HPCSVM).This new method is expected to not only dig out the potential high-order relationship between the training samples,but also utilize the prior discriminative information about each constrained pair.Therefore,this modified model can improve the performance of SVM.4.In the conventional SVM+ under learning with structured data,the grouping method has great randomness and only considers part of the structural information of the training dataset.By combining feature selection and clustering technique,we propose a novel framework termed as feature selection and clustering technique-based SVM+(FCSVM+),and then design an iterative algorithm to implement it.Concretely,the new method first acquires some feature attributes by adopting feature selection method,and then performs clustering technique on the chosen feature attributes to obtain group information,finally,SVM+ model is adopted to obtain the final decision function.The comprehensive experimental results on the UCI benchmark datasets verify the rationality and validity of our proposed approach.
Keywords/Search Tags:Support vector machine, Nonparallel hyperplane support vector machine, Structural information, Discriminative information, Linear discriminant analysis
PDF Full Text Request
Related items