Support Vector Machine Model And Several Relative Theory Research

The vector support machine (SVM) technology is a new model which is proposed by V. Vapnik in the mid 90s, to deal with the non-linear classification, regression and other machine learning problem. The theory research develops rapidly, and more fields pay attention to the actual application. Traditional classification method, which is from the induction to deduction, faces problems of low efficiency and predictive accuracy when deal with multi-dimensional nonlinear condition. While the SVM simplifies the classification process by adopting transduction inference between the training data and test data, only a few parameters are needed to determine the decision function in SVM model, and other parameters can be chosen based on experience. Time complexity especially spatial degree, which compared to the previous classification ways, depending not on the attribute dimension, but on the number of support vectors, shortens the classification time and reduces storage space.1. Relative to the defect of fuzzy membership as a function of distance between the point and its class center in feature space for some current fuzzy support vector machines, a new FSVM which based on entropy and ACO (EAFSVM) is first proposed. Exploiting advantage of evaluation of entropy and intelligence of ant colony optimization (ACO), EAFSVM enhances the classification capability and makes clustering center more suitable and membership more accurate. Experiment shows that the EAFSVM has a better precision and classification performance, especially to multi-class and large scale dates.2. The effectiveness of SVM model depends on accuracy of acquiring the information of data. Considering the low predictive accuracy and poor generalization capacity of SVM caused by simply mining data, a new model is proposed. Combining with probability distribution and equivalence class information among data, this model optimizes the traditional SVM by adopting double coefficients to promote the capability of acquiring information. Each instance will be assigned two coefficients of probability value and equivalence class. The experiment on comparing with SVM, RSVM and FSVM illustrates the new model not only can utilize information effectively but also assure a remarkable predictive accuracy and classification capability, and be more robust.