Study And Application Of Support Vector Machine Based On Pulse Kernel

Data mining as an important area of information management field, the technology has been widely applied to many industries. Support Vector Machine (SVM) which is a new data mining technique is presented in the 1990s. As for structural risk minimization principle, SVM has a better solution to small sample, nonlinear, high-dimensional learning problems. As a hot study area, although SVM has formed a theoretical framework, its performance needs to be improved in addressing some practical problems. So the deep research on its theory and application still continues. In order to improving SVM's generalization and learning speed, many ways such as improving algorithm, providing new algorithm, constructing new kernel and choosing suitable parameters are now researching.After introducing some basic SVM knowledge in this thesis, more deep research to the performance of SVM is done. The main job is as follow:1. Under the situation that the Mercer kernels used in SVM are a few, a new method is provided to construction more kernels used Fourier transform theory. We can proof that new kernel which is constructed by this method also meet Mercer condition. Then a new pulse kernel is constructed and its property is analyzed. Experiment to UCI data sets shows result that the number of support vectors got from SVM based on the new Mercer kernel is less than got from SVM based on RBF kernel, but the generalization is almost unchanged.2. Less number of support vectors got from SVM training based on the pulse Mercer kernel, this can improve classify speed. But if the training set is larger, the learning speed of SVM will be slower. In order to overcoming this drawback, a simple SVM model based on guard vector is provided. With unchanged forecast accuracy, this model is more suitable to solving large scale problem.3. As consumer credit warming in recent years, credit institutions need to conduct a borrower's credit assessment as accurately as possible. Some existing methods can do this, but data distribution is restricted. Different with this, SVM can solve problem without any request to data distribution, so former modified SVM model is used to the personal credit evaluation. It provides a new way to develop evaluation technique and also it provides reference to credit institutions when they are decision making.