Ensemble Of SVDDs Based On Information Theoretic Learning

One-class classification is regarded as a machine learning task between supervised learning and unsupervised learning. It can effectively solve the problem of training with samples in only one class and the problem of extreme class imbalance. Till now, a large number of one-class classification methods have been proposed. The two commonly used one-class classifiers are one-class support vector machine(OCSVM) and support vector data description(SVDD). In order to further improve the performance of one-class classifier, a number of one-class classifiers can be integrated. However, the existing one-class classifier ensembles have not considered the influences of diversity measure and the selective ensemble simultaneously. Therefore, basing on the existing one-class classification ensembles and information theoretic learning, the information theoretic learning based ensembles of SVDDs are studied in the dissertation.1.Selective ensemble of SVDDs based on correntropy and distance variance is proposed.Correntropy is utilized to substitute mean square error to measure the compactness of ensemble and construct more compact classification boundary. The variance of the distance between the training samples and the center of minimum enclosing sphere is used as the diversity measure of the ensemble model. An1?-norm based regularization term is introduced into the objective function to implement the selective ensemble. Moreover, the half-quadratic optimization technique is utilized to solve the proposed selective ensemble model. The experimental results show that the proposed method achieves better classification performance in comparison with its related methods.2.Selective ensemble of SVDDs based on correntropy and Renyi entropy is proposed.The Renyi entropy of distance between the samples and center of ensemble model is utilized as diversity measure in the proposed ensemble method. Meanwhile, the half-quadratic optimization technique is used to solve the problem. The result of experiments on synthetic data sets and benchmark data sets shows that the proposed method is more robust.