Regularized Angular Margin Core Vector Machine

Although pattern classification has been extensively applied, how to effectively solve the corresponding training on large datasets is a problem, especially how to training on unbalanced large datasets. Many kernelized classification methods can be formulated as the corresponding quadratic programming problem, but computing the associated kernel matrices requires very high computational complexity, when the size of the training datasets is large.To address imbalanced datasets and large datasets two issues, we proposed Regularized Angular Margin Core Vector Machine(RAMCVM), which mainly includes the following two aspects:1、For the classification of the large datasets, we proposed Position Regularized Core Vector Machine(PRCVM). The main idea is that use the minimum enclosing ball to find the core set, based on the different position in the feature space, given different weight, then we have the optional minimum enclosing ball, the performance can be improved.2. For the problem of imbalanced data classification, we propose a new classification algorithm, called the Central Vector-Angular Margin Classifier(CAMC). This algorithm is to find the optimal vector in the feature space, then use the optimal vector to decision. Accordingly, it is proved that the kernelized CAMC can be equivalently formulated as the kernelized Center-Constrained Minimum Enclosed Ball(CC-MEB). Then it may be extended a Regularized Angular Margin Core Vector Machine(RAMCVM) by connecting the PRCVMThe two methods are used in UCI data set, and it is compared with other methods respectively. The experimental results show that the PRCVM and RAMCVM have better classification performance. The CAMC is as accurate as existing Support Vector Machine(SVM), even more accurate. The RAMCVM is much faster and can handle large imbalanced data sets.