Ensemble Of One-Class Support Vector Machines

Novelty detection is considered as one-class classification in the fields of machinelearning and pattern recognition. The novelty detection models are trained only with thenormal data but used for classifying a test sample as the normal data or the novel data. So far,there have been a lot of one-class classifiers, among which one-class support vector machineand support vector data description are mostly used. The cross-validation method is usuallyused for choosing parameters for these one-class classifiers. If the selected parameters are notappropriate, the obtained one-class classifier cannot effectively model the distribution of thegiven normal data, which results the incompact classification boundary. In order to enhancethe performance of single one-class classifiers, several one-class classifiers can be integratedwith certain rules to make the obtained classifier efficiently model the distribution of thenormal data and produce the compact classification boundary.AdaBoost is a common method for combining classifiers. One-class support vectormachine is a strong classifier, which makes the result of ensemble not good as the baseclassifiers of AdaBoost. Hence, the AdaBoost is modified in the dissertation to make it fit forcombining one-class support vector machine. Moreover, the selective ensemble of supportvector data descriptions is proposed. This method first uses cross correntropy andauto-correntropy to respectively replace the mean square error and variance, and establishesthe corresponding model of weighted optimization problem. Then the optimal weight vector isobtained through the half-quadratic optimization technique. On the premise of ensuring theaccuracy rate, the proposed method can get rid of the unnecessary base classifiers andimprove the performance.Experimental results in the dissertation demonstrate that the first method (the modifiedAdaBoost based one-class support vector machine ensemble) can improve the classificationperformance of a single one-class support vector machine, while the second method (theselective ensemble of support vector data descriptions) can effectively reduce the number ofbase classifiers and ensure the accuracy is not less than or even higher than the ensemble ofall the base classifiers.