Twinhypersphere Support Vector Data Description And Non-parallel Hyperplanes Ordinal Regression Machine

Supervised learning methods are the most common machine learning methods.They are widely used in spam detection,pattern detection,natural language pro-cessing,sentiment analysis,automatic image classification and other real problems.Classification problems and regression problems are the two most important type-s of supervision learning.The problem of ordinal regression is a bridge between the two.It has a wide range of applications in information retrieval,collaborative filtering,medical psychology,etc.Support vector machine can maximize the clas-sification interval while minimizing generalization errors,which is a powerful tool to solve the classification problem and regression problem.Based on the support vector machine,this paper proposes methods to solve the three-class classification problem and ordinal regression problem.The specific contents are as follows:First,a method to solve the three-classes classification problem is proposed,that is,twinhypersphere support vector data description(TH-SVDD).The main idea is:when there are three classes in the target data set,the information of patterns in the third class are used to establish optimization models for the first and the second class respectively.The two models make one of the optimal hypersphers contain as many patterns in the first class as possible and reject patterns in the second and the third class,meanwhile,the other contain as many patterns in the second class as possible and reject patterns in the first and the third class.Consequently,three-class patterns can be classified.By solving two smaller optimization problems,this method greatly reduces the computational complexity and calculation time.Through selecting appropriate kernel parameters for different classes,TH-SVDD can improve classification accuracy.In the experiments,TH-SVDD is tested with manual data and UCI data respectively,and then compared with some classical multi-class classification methods in the prediction accuracy.The results show that TH-SVDD is effective and feasible.Second,a method for solving the ordinal regression problem,non-parallel hy-perplanes ordinal regression machine(NPSVOR),is proposed.This method finds K different hyperplanes for the K classes containing the ordinal information,so that each class is as close as possible to the corresponding hyperplane,achieving the purpose of classifying K classes containing ordinal information.Using the order information of the samples,when looking for each hyperplane,only two adjacent samples are used as constraints.The method only needs to solve K small-scale quadratic programming problems and reduces the running time.The method is val-idated on different data sets and compared with other methods.The results show that the method is efficient.