Some Studies Of SVM Model Improvement

Support Vector Machine (SVM) is a machine learning method based on statistical learningand optimization theory. Because it resolves the small sample, nonlinearity, high dimension andlocal minimum problems perfectly, it drums up the development of machine learning and hasbecome a hot field of machine learning. But there are some disadvantages about SVM. TraditionSVM, which doesn’t take into account of the influence of different input samples on theacquirement of the optimal hyperplane, is very sensitive to noises and outliers in the trainingsamples. In addition, model selection is a theory drawback of SVM, which includes the selectionof kernel function and kernel function’s parameters.This dissertation summarizes the study of the sensitive to noises and outliers in the trainingsamples and the inflexibility of model selection, and then conducts a deep research on this topic,the main results are as follows:(1) Geometry average membership function based on the cluster center and the closeness isproposed, which is depicted as the affinity between them and effectively reduces the influence ofnoises and outliers on the optimal hyperplane;(2) Taylor-kernel with moderate decreasing (T-KMOD) function is proposed. T-KMOD cansolve some practical problems more flexibility and better, whose classification accuracy andflexibility are testified by simulation experiment;(3) The mixed kernel function based on the polynomial and T-KMOD kernel function isproposed, which has advantages of the two kernel functions. The performance of the mixedkernel function is better, whose classification accuracy and flexibility are testified by simulationexperiment.