Research Of Multiple Kernel One-class Support Vector Machine

Anomaly detection is an urgent problem in real life,which has become a research hotspot in the field of machine learning and deep learning.The essence of anomaly detection is one-class classification problem with extremely imbalanced class distribution.At present,the research on one-class classification methods has presented a phenomenon of a hundred schools of thought contend.Among them,the method based on the support vector is widely favored by scholars,i.e.,one-class support vector machine(OCSVM)and support vector data description(SVDD).This method mainly constructs an optimal boundary around the target class by learning the target data,and achieves the purpose of identifying abnormal data with high precision.With the wide application of this kind of method,the disadvantages of traditional one-class support vector machine gradually appear.Firstly,the kernel function and its parameters used in the algorithm are difficult to determine,and there is no general method to determine,but the performance of the algorithm largely depends on the kernel function and its parameters,so the selection of kernel function and its parameters in one-class support vector machine is a core problem.Secondly,the algorithm is very sensitive to outliers or noises in the training set and has poor robustness.Based on the above problems,this paper conducts research on one-class support vector machine.The following are the main work and achievements of this article.1.Combining multiple kernel learning with one-class support vector machine,a multiple kernel one-class support vector machine based on kernel-target alignment is proposed.This method replaces the single kernel function in the traditional model with a multiple kernel function,and calculates the weights of each basic kernel by optimizing the kernel alignment model,and then constructs the required linear weighted synthetic kernel,avoiding the selection of the optimal kernel function and its parameter.The experimental results show that the classification performance of this method is better than the traditional cross-validation parameter selection method,and the training efficiency is relatively improved.2.Fuzzy membership is introduced into multiple kernel one-class support vector machine,and kernel-target alignment based fuzzy multiple kernel one-class support vector machine is proposed.This method makes use of the different types of certainty of the samples in the training set,and assigns a fuzzy membership degree to each sample to reduce the negative impact of noise or outliers on the classification boundary.At the same time,a multiple kernel model based on kernel-target alignment is used to replace the single kernel.Experimental comparisons with other two algorithms on artificial and standard datasets show that the proposed method is robust against noise,improves the robustness of one-class support vector machine,and has high computational efficiency.3.Based on the local density of data,a new method to calculate the membership degree of samples is proposed.In this method,the kernel function is used to measure the local similarity among samples,and the local density information of samples is determined.Then different membership degrees are given to samples according to the local density.Compared with other membership methods,only target data is needed in the calculation of local density based membership,which is consistent with the training set of one-class support vector machine.The membership calculation method is combined with the fuzzy multiple kernel one-class support vector machine model,and the simulation experiment is carried out on the standard data set.Compared with the multiple kernel one-class support vector machine and the fuzzy multiple kernel one-class support vector machine based on kernel-target alignment,the method has better anti-noise ability.