Rotation Based And Improved AdaBoost Based Ensembles Of One-class Support Vector Machines

Classification problem is a common task in the fields of machine learning and pattern recognition.However,when the classes of training samples are extremely imbalanced,the commonly used classifiers may produce poor classification results.To solve the above problems,one-class classification method is emerged.One-class support vector machine(OCSVM)is regarded as a commonly used one-class classification method.OCSVM is considered to be a strong classifier,so combining several OCSVMs by the traditional ensemble methods cannot improve the performance of a single OCSVM.To improve the classification performance of the ensemble of OCSVMs,two improved ensemble methods for combining OCSVMs are proposed.First,a rotation based ensemble method is proposed for integrating OCSVMs.For different OCSVMs in the proposed ensemble,their training samples are transformed by different rotation matrices.Therefore,the diversity of training samples for the ensemble of OCSVMs can be guaranteed.Experiments conducted on the ten benchmark data sets validate the effectiveness of the proposed ensemble strategy.Second,an improved AdaBoost based ensemble method is proposed for integrating OCSVMs.The traditional ensemble method used to combine multiple OCSVM may not produce better performance.In this thesis,the loss function of the conventional AdaBoost is redesigned,i.e.,substituting the exponential loss function by a more robust one to enhance the robustness of the traditional AdaBoost based ensemble of OCSVM.The proposed loss function is defined as the weighted combination of the modified exponential loss function and the squared loss function.The robust AdaBoost based on the proposed loss function is introduced by redesigning the update formulae for the weights of base classifiers and the probability distribution of training samples.The upper bounds of empirical error and generalization error for the robust AdaBoost based ensemble of OCSVM are derived.Experimental results on the synthetic and benchmark data sets demonstrate that the proposed ensemble method is superior to its related approaches.