Expansion And Research Of Multicategory Cost Sensitive Support Vector Machines

With the development of the data age,more attention has been paid to the cost sensitive multi-category problem.Existing classification algorithms can be roughly divided into two types.One is to combine multiple binary classifiers.However,this kind of method ignores the connection between different classes.The other is the multi-classification support vector machine,which gets the solution of parameters by one optimization problem.Compared with the former,the complexity of parameter solution is improved,but it can better grasp the distribution of data categories.Multi-category cost sensitive support vector machine(mcSVM)is a representative of the second kind of algorithm for cost sensitive problems.In this paper,the regularization term of the original model is extended.The objective function is transformed into a quadratic programming problem by introducing parameters,the influence of regular terms on the model is also studied and the model parameters are solved by using efficient block coordinate descent method.In addition,for large-scale samples,the high-efficiency calculation method is given,experimental results verifies the advantages of the new model in general classification and cost sensitive learning.Besides,this paper expands the existing Rademacher structural lemma to calculate the generalization error boundary of the multi-classification model,studying and comparing the result with Gaussian boundary.Finally,the generalization error boundary of the extended multi-class cost sensitive support vector machine is given,and the relationship between the generalization boundary,class number of the classification problem and the model regularization term is studied,which is proved by numerical simulation.