Research And Application Of Local Support Vector Machine In Classification

Support vector machine(SVM) is a kind of supervised machine learning algorithms which was based on the statistical learning theory. SVM has many advantages for small sample problems because of the structural risk minimization principle. SVM can effectively solve the nonlinear problem and dimension disaster by utilizing kernel function. However, the global behavior of SVM does not contain the consistency and is not able to deal with the convex sets of data effectively.Local support vector machine is proposed by introducing the local learning algorithm which can make full use of the local information of the sample, and satisfies the requirement of consistency of the algorithm. In this paper, the local support vector machine(SVM-KNN) is mainly studied, and some improved algorithms are proposed in solving the problems that exist in the local support vector machine in classification. The main work of this paper is as follows:(1) A fast local support vector machine algorithm is proposed based on clustering, the algorithm has higher classification accuracy and efficiency than SVM-KNN.According to the problem of low classification efficiency of local support vector machine, FKNN-SVM is proposed by improving the efficiency of choosing the k nearest neighbor sample of the unclassified sample. To verify the efficiency of FKNN-SVM algorithm for classification, we take some comparative experiments on maize images for FKNN-SVM, SVM-KNN and SVM algorithm. The experimental results show that FKNNSVM algorithm performs better than the SVM-KNN in classification efficiency.In order to further improve the efficiency and accuracy of classification for FKNN-SVM algorithm, CFKNN-SVM is proposed by combining the clustering algorithm with FKNNSVM algorithm. To verify the validity of the CFKNN-SVM algorithm, we take some experiments in the UCI data sets and bark image data sets. The test results show that the classification accuracy and efficiency of the CFKNN-SVM algorithm is superior to FKNNSVM and SVM-KNN algorithm.(2) A kind of local support vector machine algorithm is put forward which is suitable for unbalanced data. This algorithm can improve the classification accuracy of positive class.Unbalanced data exists in the intrusion detection, medical detection and other fields. In unbalanced data sets classification problems, the positive class with less data often has more important value. A kind of local support vector machine algorithm(CLSVM) which is suitable for unbalanced data is proposed by clustering algorithm. The algorithm can keep sample information as much as possible, and it can improve the classification accuracy of positive class with the aid of clustering algorithm to solve the homomorphic imbalance problem of the k nearest neighbor sample of the unclassified sample.(3) A new local support vector machine algorithm is proposed based on neighborhood kernel function, the algorithm can use unstructured image data to build a classification model directly.Image classification has been a hotspot of machine learning, and local support vector machine cannot use unstructured information such as the image directly. In this article, a Neighborhood-LSVM is proposed by utilizing neighborhood kernel function in local support vector machine. Neighborhood-LSVM algorithm utilizes the change of image pixel's Neighborhood information directly to build the image classification model. The results of the bark image data sets experiments show that the accuracy of Neighborhood- LSVM algorithm is better than SVM-KNN and SVM in image classification.