An Ensembling One-class Classification Method Based on Beta Process Max-margin One-class Classifier

In the problem of one-class classification, One-Class Classifier (OCC) tries to identify samples of a specific class, called the target class, among samples of all other classes. Traditional one-class classification methods design a classifier using all training samples and ignore the underlying structure of the data, thus their classification performance will be seriously degraded when dealing with complex distributed data. To overcome this problem, an ensembling one-class classification method based on Beta process max-margin one-class classifier is proposed in this paper. In the method, the input data is partitioned into several clusters with the Dirichlet Process Mixture (DPM), and a Beta Process Max-Margin One-Class Classifier (BPMMOCC) is learned in each cluster. With the ensemble of some simple classifiers, the complex nonlinear classification can be implemented to enhance the classification performance. Specifically, the DPM and BPMMOCC are jointly learned in a unified Bayesian frame to guarantee the separability in each cluster. Moreover, in BPMMOCC, a feature selection factor, which obeys the prior distribution of Beta process, is added to reduce feature redundancy and improve classification results. Experimental results based on synthetic data, benchmark datasets and Synthetic Aperture Radar (SAR) real data demonstrate the effectiveness of the proposed method.