Research On Local Support Vector Machine

Statistical learning theory (SLT) is built on the basis of structural risk minimization principle. Compared with the traditional statistical learning theory, statistical learning theory proposed by V.Vapnik is presented to exploit the rule of statistical learning for small samples. Support vector machine (SVM) which is developed based on SLT is often used for pattern recognition and regression estimation. SVM is better than other machine learning algorithm in performance because it is based on the completed SLT, its optimization problem is a global optimization problem, and it is the one of best methods in gereralization performance. Global optimization of support vector machine method, however, does not imply consistency theory. The local support vector machine (LSVM) which is proposed in recent years is consistent with the idea of "consistency implies locality". In recent years, LSVM has received concerns in both theoretical research and application scope, such as remote sensing image classification, the EEG signal processing and network traffic prediction.Firstly, the paper gives a simple introduction of machine learning and SLT: furthermore, the characteristics and research of SVM are introduced. Secondly, it expounds the current situation and significance of LSVM, details the features and advantage on learning performance and some improved strategies, and finally analyses several current algorithm and their improved ideas of LSVM.This paper completes the following work:(1) Because of the high complexity of large-scale data sets in LSVM, the paper proposes a LSVM based on cooperative clustering (C(?)LSVM), and reduces the time complexity successfully without reducing the classification accuracy.(2) Detailing the weighted support vector machine (WSVM) and applying the weighted thought into Falk-SVM. Proposing weighted Falk-SVM (WFalk-SVM), it is used to improve the accuracy of the classification of small sample data sets.(3) Detailing one-class support vector machine (OCSVM), applying it to Falk-SVM, and proposing OCFalk-SVM for an exploratory study.