The Fusion Of Support Vector Machine And Neural Network And Its Application In Data Mining

Undoubtedly,big data is one of the most concerned focuses in current society.Big data algorithms provide strong support and great contributions to cloud computing,artificial intelligence,smart city and other aspects in modern industry.However,the horizontal researches of big data algorithms are not enough from scholars at home and abroad.It will restrict the breadth of algorithms in a way.Therefore,taking Support Vector Machine(SVM)and Radial Basis Function Neural Network(RBFNN)as objects,this thesis explores the relationship between SVM and RBFNN deeply and proposes an optimization model based on the existing RBFNN.The main contents are as follows:Firstly,we adopt the method of traceability to discuss their design philosophies,core parameters and optimization strategies,respectively.This part not only facilitates the discovery of the essential characteristics between two algorithms,but also establishes a solid theoretical foundation for analyzing their relationships and fusing a new model in the next part.Secondly,cooperated with the kernel factors of support vector machines and RBF neural network by means of regularization theory,this thesis elaborates their relationships in three aspects,that is,the structure ideologies of networks,the addition of constraint conditions and the derivation of formulas.These works construct a strong theoretical framework to present a SRmix model(SRmix is short for the mixture of SVM and RBFNN).Thirdly,this thesis puts forward the SRmix model.Compared with the existing RBFNN,the SRmix algorithm has made the following improvements.1.We cluster support vectors in the SVM to obtain more accurate centers in hidden layer of network.Meanwhile,it can effectively avoid overfitting phenomenon of RBFNN.2.Due to the strong generalization ability of the algorithms being presented,the weights adjustment strategy in network is more important than its assignment strategy,even abandoning some detailed assignment strategy in a sense.3.The extended constants in RBFNN are replaced by relevant parameters in SVM with RBF kernel function.At the same time,a bidirectional search is performed,which is helpful to improve the efficiency of parameters optimization.4.According to the stop conditions of procedure,neurons are added to the hidden layer of RBFNN dynamically and computing efficiency of the network are improved gradually.This treatment can solve the inadequate or overburdened network operation problems effectively.Experiments are designed to demonstrate performances of the SRmix model.The tests are divided into two parts.Firstly,a series of comparative experiments are designed on different dimension data sets from UCI benchmark database.Secondly,the SRmix model is applied to solve a motion recognition problem.Experimental results show that SRmix model has stronger algorithm stability and generalization ability than BP network and SVM.