Some Reasearches On The Algorithm Of Support Vector Machine Classification

Support Vector Machine(SVM) is a machine learning method with good performance,whichdevelopes based on the statistical learning theory,it achieves great success in many domains,such asdata ming,pattern recognition and the signal processing etc,so it is of great significance to researchthe performance of support vector machine.There are many factors that can affect the performanceof support vector machine,such as the construction of kernel function,the existence of noise andoutliers,and imbalanced datasets etc,these problems are the hotspots in current study of supportvector machine.In this paper,we mainly study the construction of the kernel function in support vectormachine,the improvement algorithm of membership function in fuzzy support vector machine,andthe application of support vector machine in imbalanced datasets.The main innovation works are:（1）A kernel scaling method based on Riemannian geometry is proposed.The methodmodified kernel function by constructing a conformal mapping factors in the form of atrigonometric function and using the distance between training points and the hyperplane(Trigonometric Kernel Scaling,TKS).The TKS method not only provides a new form of conformalmapping function, which can enriche the construction of the kernel function methods from theperspective of the geometry, and it also can improve the classification accuracy of SVM.（2）The membership function based on the distance is changed.In fuzzy support vectormachine,the method introduces different coefficients to membership function according to thedifferent importance of samples to the classification hyperplane in different areas.Experimentsverify that the revised membership functions can effectively distinguish between effective samplesand noise points，and improve SVM performance.（3）A new algorithm based on the undersampling combined with Cost Sensitive SupportVector Machine is proposed.Firstly,a new undersampling method is presented to preprocess theimbalanced datasets,in order to obtain better result,then cost sensitive support vector machine isused to train pre-processed datasets.Experiments show that this algorithm can effectively improvethe classification performance of SVM in the imbalanced datasets.