Support Vector Machines For Classification And Its Application

Support Vector Machines (SVMs) are new developed machine learning methods based on the foundations of Statistical Learning Theory (SLT) and optimization theory. They are proposed firstly by Cortes and Vapnik in 1995 and have become an important research direction in the field of machine learning. This paper researches on support vector classification (SVC) in theory and model in order to solve classification problem. The following parts are main works:1. Deal with classification problem as special regression problem based on the theory of SVC and support vector regression (SVR). By introducing different norms and different loss functions, a new SVR model solving classification problem is constructed. For the case of Gauss loss function implied, a simple solving method - concise sequential minimal optimization (SMO) algorithm is proposed. Furthermore, for multi-class problem a new classification SVR algorithm - K-SVR is constructed, and experiments show the model's robustness and efficiency. Therefore we give out new idea and method of solving classification problem.2. For the proximal support vector machine for classification (PSVMC) proposed intuitively by Fung and Mangasarian, we get its primal problem by theoretical deduction so that construct the PSVMC by different way. And on this base, we propose sparse PSVMC and weighted PSVMC first time. For the classification problem with uncertain information, we construct uncertain PSVMC by introducing probability invariable. Therefore we generalize and develop the theory and algorithm of PSVMC.3. Vapnik proposed the transductive support vector machine (TSVM), the optimization problem in it is difficult to solve because of its special property. We construct improved TSVM by transform the constrained problem in TSVM to unconstrained problem and smooth it Furthermore, weighted TSVM is also proposed in order to solve real problem with unbalanced training points. The new model constructed is applied into detection of Web intrusion successfully. Therefore, the theory and application of TSVM are both improved.4. Apply SVC into environment supervise problem of Seawater industrialization breed aquatics first time. After collect living environment data of fish randomly in industries of Tangshan city and Qinghuangdao city of Hebei province, we do lots of detecting and monitoring experiments by SVC and get the corresponding results, which show that the application is successful and meaningful. Therefore, with solving the real world problem, we extend the field of SVM applications.The new SVM models constructed in this paper have their obvious priority than standard SVM, some has more simple solving method, some has more high testing precision, and some deals with real world problem effectively, all of which are proved in experiments and real applications.