Research On Least Squares Support Vector Machines In Multi-classification

Classification problem is the core content in the field of pattern recognition. At present,many intelligent methods such as neural network, decision tree method and Support VectorMachines (SVM) have been widely used in the classification problem. As an improved modelof the SVM, Least Squares Support Vector Machines (LS-SVM) has many advantages suchas global optimum, good generalization ability and great adaptability. So far, LS-SVM hasbeen widely applied in pattern recognition, signal processing, and many other engineeringfields, and it shows good performance.Based on the analysis of the theory of SVM and LS-SVM, this thesis focuses on twokinds of improved algorithms: Sparse Least Squares Support Vector Machines (SLS-SVM)and Fuzzy Least Squares Support Vector Machines (FLS-SVM). The availability andfeasibility of the two improved algorithms are verified by applying them into several kinds ofclassification problems, including artificial datasets, benchmark datasets, electrocardiogram(ECG) signals and electroencephalogram (EEG) signals.The main contents of this thesis are as follows:(1) Based on the theory of the standard SVM and LS-SVM, two improved algorithms arestudied: SLS-SVM and FLS-SVM. In addition, this thesis researches on the influence of theclassification results by using different kernel functions and parameters, the cross validation(CV) method is used to optimize the parameters of the classifier and kernel function.(2) SLS-SVM and FLS-SVM based on wavelet kernel function are used in theclassification problems of binary-class artificial datasets, binary-class benchmark datasets andmulti-class benchmark datasets respectively. Compare the classification performance of thetwo improved classifiers with the existing ones based on radial basis function, multi-layerperceptron kernel function and polynomial kernel function under the same conditions. Andthen a detail analysis of each classifier’s classification performance is given from theclassification accuracy and running time. The results show that SLS-SVM based on waveletkernel function can improve the training speed without reducing the classification accuracy;FLS-SVM based on wavelet kernel function can improve an problem of the inseparable datain multi-classification problems. Finally the two improved classifiers are proved to be validand feasible.(3) For actual classification problem in biomedicine background, the two improvedclassifiers are used for the ECG and EEG signals classifcation problems. The results showthat the two improved classifiers based on wavelet kernel function have good classificationperformance and generalization ability in dealing with the actual classification problems.