Analysis Of Parameters In SVM, SVR, And Its Application In Dealing With DNA Sequences

Our research group have proposed an information-geometrical method to modify the polynomial kernel function and the Bn - splines kernel function. In this papers, we analyze the parameters in the method. We determine riot only the region of the paramaters but the relation of the them, so the parameters can be choosen accurately. With the good choice of parameters, we can modify the volume element near SVs to improve the precision of SVM and to compress the data.Based on the results of the analysis, we do the corresponding numerical experiments of the two kernels. For the SVM classifier, we use the complete sequence of arabidopsis thaliana to research the coding region and the noncoding region in DNA sequences. In the experiment, we first get all window sequences and the relative difference a which reflects the frequency difference of window sequences. The feature sequences of coding region and noncoding region can be determined based on the a, then we can express the DNA sequences to feature vectors. At last, we use the primal SVM and the modified SVM classifier to classify all feature vectors. The resluts show that the performance of the modified SVM is improved when paramaters satisfy the given conditions. For the SVR, we do experiments of the function approach using artifical and real data. Simulation results show the improvement of gerneralization errors and the number of SVs.