The Research On Mistakenly Classified Instances Based On Kernel Funcation

The statistical learning theory synthesized the machine learning, the statistical study, and the neural network technologies and so on. Through using the structural risk minimum principle and the empirical risk minimum principle, statistical learning theory effectively enhanced the algorithm generalization, and provided the powerful theoretical foundation for the small sample situation of machine learning.Based on the statistical learning theory and the kernel technology, VaPnik and other scholars presented a new machine learning method --Support Vector Machines (SVM) in the 1990s. Basing on the structural risk minimum principle, the core idea of this method is that the low-dimensional space linearly inseparable issues mapped into a high dimensional space linearly separable issues by introducing the kernel function techniques, it can be better to solve non-linear, high-dimensional identification, small samples, local minimum points and other issues. The development of SVM had not only enriched and developed the statistical theory, but also promoted in many application fields, such as text categorization, handwriting recognition, face recognition, Web mining, regression analysis. How to further improve the performance of Support Vector Machines is always attention and research hotspot of the field of pattern recognition and machine learning.Although the kernel function's support vector machines had already achieved good results in classification and regression, etc. however, they often only paid attention to the correct classified data information, and actually neglected the mis-classification data information, at the same time, the classified effect of the kernel function's support vector machines depended highly on the kernel function and the choice of kernel parameters. Therefore, it is a good practical significance that used the kernel functions and excavated the useful information of the mistakenly classified instances in order to improve the classifier's classification and predictive ability.Basing on the kernel function's support vector machines, this paper took advantage of the useful information of the mistakenly classified instances and completed the following work:1. Introducing and studying the Support Vector Machine theory and algorithms, briefly introducing the development of kernel functions history, theoretical foundation and basic ideas of kernel functions; introducing the basic knowledge of the machine learning and statistical learning theory.2. This paper proposed a kind of classification method(Support Vector Machine Classification based on Perceptron,PSVM), which based on perceptron, the model in the classifier of training, has introduced the perceptron study thought, first, it uses the support vector machines kernel function, and then kernel calculation, judges the classification of performance, if they were classified correctly, it would make no revision, on the contrary, transformed for the perceptron study question. Experiments show that the model can not only improve the performance of SVM classification, but also can reduce the SVM classification performance to the kernel function and parameters'choice dependence.3. Basing on the kernel function's Support Vector Machine and using data processing operation among difference evolution algorithm, a kind of classification method (Support Vector Machine classification based on differential evolution (DSVM)) was proposed in this paper. Using the support vector of the Support Vector Machine, and combing with the data variation and cross-operation in the differential evolution algorithm, the model processed the mistakenly classified instances and used the useful information of mistakenly classified instances to enhance the classified effect. This model increased the population individual multiplicity and speeded up the objective function convergence rate.