A New Dimension Reduction Method Based On SVM And Its Application To Three-class ROC Method

With the rapid development of artificial intelligence、big data and so on, machine learning come to its spring of development. Machine learning is a cross subject in many areas and the core of artificial intelligence research field, whether in industry or in academia,.Machine learning theory main purpose is to design some algorithm to make computer have the ability to self learning, through the training set for model, and use this model to forecast the unknown sample training set, in many areas has reached even surpasses the level of human beings.In traditional machine learning techniques, and is the hypothesis that the sample has the same class distribution and the type of distribution is balanced, the other two types of errors cost under the premise of the same, and the practical application, the distribution of the data is often unbalanced in wrong price is also there is a difference.In order to promote the further development of machine learning technology, we must review the limitations in the traditional machine learning, and the advantages and disadvantages of measure of machine learning algorithms is our priority direction, namely ROC analysis technology.The ROC analysis technology has been widely used in biomedical, signal processing, machine learning, and other fields, but is mainly used in the two types of problems, in multi-class problems, ROC is faced with problems such as high dimensional space, said difficult, difficult to understand.The main goal of this paper is proposed a new method for classification of three classifier performance comparison, using ROC curve to compare to the volume of the three classification performance of the classifier, effectively expand the ROC analysis technology in the application of three-calss problem.This article needs to solve the main problem is to realize the data dimension reduction, formation of ROC curve and the size of the volume under ROC curve is obtained. Tool based on support vector machine (SVM) classification, adopt "one-against-rest" classification of three-calss data classification, get a three dimensional numerical, and then to three classification results in order to establish the space coordinate axis, then on the basis of the data dimension reduction, the projection of data to a three dimensional space first plane, then the data on the three dimensional space plane mapped to two dimensional space. Steps by right Angle bracket to traverse the sample points on the two-dimensional space, support to each position will have a 3D point, all of these points will constitute the three-dimensional surface of ROC.Finally using block volume volume under the ROC curve obtained by the way, the ROC curve is divided into several small triangular prism, each a small triangular prism and can be divided into three small triangular pyramid.Finally, in this article, through the simulation experiment data sets and real data set of experiments, the results fully demonstrate the reliability and superiority of this method compared with traditional three-class of ROC analysis technology, the method avoids the higher dimensional space, and is easy to express, the advantages of easy to understand, compared with the nonparametric method, volume method used in this article VUS possesses the advantages of low computational complexity.