Researches On Granular Support Vector Machine Learning Approach Based On Multi-dimension Association Rules

Support vector machine is an effective machine learning approach, which can deal with the problems in solving small samples, nonlinear and high dimension pattern recognition, et al.SVM has become the hot spot in the field of machine learning presently, and is applied to pattern recognition, function estimation and time series prediction data mining problems. However, it has also certain limitations in practical applications, such as, its training speed is greatly affected by the size of the given training data sets; it can ensure high training accurate rate but may lead to over fitting; In particular space, its generalization ability is often restricted.The thesis will focus to introduce the granular computing into SVM so as to overcome the above disadvantages of the traditional SVM.Within the framework of granular computing theory and support vector machine, granular SVM(GSVM) learning approach is proposed based on multi-dimension association rules mining, namely, AR-GSVM. On the basis of AR-GSVM,granular SVM in kennel space is proposed, which is named as AR-KGSVM.The main works of the thesis include the following:(1)The existing granular support vector machine learning methods are analyzed systematically.(2) A granular SVM learning approach, AR-GSVM, based on mining Multi-dimension association rules is proposed.It can not only reduce the complexity of the classifier but also improve the learning efficiency due to its embedding parallel computing.Additionally, the AR-GSVM emphasized on the data(They may be support vectors) near the classification boundary but not all the data, thus, it can further improve the generalization performance.(3)On the basis of the proposed AR-GSVM, the thesis considers data distribution inconsistency in input space and high-dimensional kennel space, which may lose much information.The AR-KGSVM is then proposed in kennel space. Firstly, it maps data from input space to kernel space, and then divides granularity in the kernel space. In so doing, it can assure granularity dividing and data training in the same space, and therefore it can obtain well generalization performance.(4) The proposed approaches are verified on standard UCI datasets,and the experiment results show that the approaches are effective and achieve the expected effect. The most important application of the proposed approaches is processing imbalanced data. Through comparing with several common used approaches, the proposed AR-GSVM and AR-KGSVM are demonstrated very efficient on benchmark datasets.The proposed granular SVM learning approaches combine the powerful SVM with granular computing theory, and the research results are not only enrich the theoretical results but also expand the application fields of SVM that is the classification problem of imbalanced dada, and related work needs to be further developed.