Font Size: a A A

Researches On Optimization Modeling Methods Of Support Vector Machine

Posted on:2015-01-14Degree:DoctorType:Dissertation
Country:ChinaCandidate:H S GuoFull Text:PDF
GTID:1108330461985137Subject:Systems Engineering
Abstract/Summary:PDF Full Text Request
Support Vector Machine (SVM) has become a kind of machine learning methods which is concerned broadly in recent years. Based on the statistical learning theory (SLT), it aims at the structured risk minimization (SRM) by kernel methods and obtains the support vectors. With solid theory basis, concise mathematical forms, standard training algorithms and good generalization performance, SVM has been widely applied in many areas such as pattern recognition, function estimation, image processing, time series prediction and bioinformatics et al. However, with the rapid development of computer network, sensor techinique and communication techinique, the expansion trend of data complexity is growing, and then the traditional SVM model can not meet the application demands of complex data analysis and processing. Therefore, how to optimize the SVM model, so as to improve the efficiency of the SVM to solve the complex data processing problem by combining with other novel data analysis and processing techniques will become the one of focus issue on SVM.The thesis integrates the SLT, SVM and relevant theories and approaches such as granular computing (GrC) theory, hierarchical structure lerning model and active learning method to optimize SVM systematically. Some SVM optimized learning methods are presented to solve the complex data processing problems such as large scale, multiple levels, multiple classes and imblanced. The main works are concluded as follows.(1) Integrates the GrC with the SVM systematically, constructs the learning mechanism and optimize models of the granular support vector machine (GSVM), and improves the learning efficiency of SVM for large scale data. Specially, by combining the granular dividing, granular computing, granular measuring with the kernel methods, the optimize GSVM methods based on kernel are proposed to sovle the inconsistency of data distribution during granular dividing and SVM training processes.In so doing, the generalization performance of GSVM will be improved greatly. Moreover, by mixed granular dividing and important mixed granules extraction, the GSVM based on mixed measure method is introduced. The classification hyperplance will be adjusted according to training results, which will lead to improve the generalization performance greatly with high learning efficiency syschronously.(2) Applies hierarchical structure model to the SVM training process and constructs the dynamical hierarchical SVM optimization model. The proposed model can improve the learning efficiency and solve the problems on various cognitive levels, i.e., it may be closer to the way that people deal with the problems in the real world. Specially, the dynamical hierarchical dividing method is introduced. It extracts the information on different granulation levels by the various cognitive levels, precision and importance. And then, the SVM classification and regression models based on dynamical hierarchical dividing are presented respectively. By textracting the important classification or regression information on different levels and deleting some unimportant information, generalization error of medel can be decreased. And at the same time, a satisfactory compromise between the training speed and the generalization performance can be obtained.(3) Presents a novel approach based on active learning to solve multiple classification unsupervised or semisupervised problems. This method extracts the most valuable samples by the sample discrepancy, and then it quickly mines pattern classes that implicated in unlabeled samples by the experts intervention. The SVM active multiple classification method is then constructed. By the pattern class mining process, the unlabeled multiple classification problems can be transformed into supervised. By the sample’s rejection, compatibility and fuzzyness, a small part of compatible samples, rejected samples and uncertain samples are selected, to extract the most valuable samples of multiple classification process effectively. In so doing, the low labeling cost, high learning efficiency and good generalization performance can be obtained syschronously. Besides, this thesis also designs the SVM active multiple classification online learning algorithm, which can improve the efficiency of SVM for online multiple classification problems. This model can be applied in the many areas such as webpages classification automatically, timely desease and epidemic detection et al.(4) Proposes SVM optimization models to solve the imbalanced data problems. Based on the introduction of the imbalanced dividing and informational granule extraction, the imblanced SVM methods are designed by sampling. It divides some granules on data belonging to majority category and extracts information granules to make the data become balanced. By this way, the ability of model for minority class is improved. On the other hand, by imblanced granular dividing, two granule factors, support and disperse, are defined to measure the influence of sample distributions on the performance of SVM. Then, the shift parameter of each granule is introduced. Based on these shift parameters, a new convex quadratic optimization problem is constructed and solved. This method can improve the obtained hyperplane which is based on maximum margin for imblanced classification problems, and it is successfully used in protein interactions prection problem.The contents of this thesis are the hot issues in the SLT and SVM study, and they are also the importance applications of granular computing theory, hierarchical structure model et al. At same time, they are the novel ways of sovling multiple classification and imblanced data mining problems. Optimization modeling of SVM for complex dataset is not only has important theoretical significance for the theory and modeling study of SVM, but also highlights the SLT and SVM application value for real complex data mining problems.
Keywords/Search Tags:support vector machine, granular computing, hierarchical structure model, multiple classification problem, imbalanced data
PDF Full Text Request
Related items