Modifying SVM With Geometric Method And Its Applications In Dealing With DNA Sequences

Support Vector Machine (SVM) is a pattern classification technique. SVM is a linear classifier in the parameter space, but it is easily extended to a nonlinear classifier by kernel functions. So the performance of SVM largely depends on the kernel. However, there is no method how to choose good kernel functions in a data-dependent way. J.C.Burges putted the geometric character to kertiels in one of his article (1999). Based on his result, Amari proposed a method of modifying a kernel function using it抯 geometric character. Which can improve the respective of SVM. Based on the method of information geometry and the thought of modifying kernel function, we proposed a geometric method of modifying the SVM. This method constructs a quasi-conformal transaction with the sech function. It improves the respective of SVM by modifying the polynomial kernel function. The text emphasizes on the parameter analysis of quasi-conformal transaction. Which finds a method of choosing parameters and reduces the range of choosing parameters. In the experiment of artificial data or DNA data, the SVM after modified has higher respective. This explains that the methods of modifying SVM and choosing parameters proposed in the text are both effective. On the other hand, along with the discovering of the genes, analyzing lot抯 of DNA data is needed. Dealing with a DNA sequence must find a mathematical notation of it first, and then analysis it by some way. Based on the Chaos Game Representation (CGR) of DNA sequences, we changing them into vectors by the view characters of CGR to denote DNA sequences. Good results for both exon-intron and exon-whole classifying experiments by SVM and Modified SVM show that the vector representation of DNA sequence are effective.