Shrinkage methods for one-class classification

Over the last decades, machine learning techniques have been an important asset for detecting nonlinear relations in data. In particular, one-class classification has been very popular in many fields, specifically in applications where the available data refer to a unique class only. In this paper, we propose a sparse approach for one-class classification problems. We define the one-class by the hypersphere enclosing the samples in the Reproducing Kernel Hilbert Space, where the center of this hypersphere depends only on a small fraction of the training dataset. The selection of the most relevant samples is achieved through shrinkage methods, namely Least Angle Regression, Least Absolute Shrinkage and Selection Operator, and Elastic Net. We modify these selection methods and adapt them for estimating the one-class center in the RKHS. We compare our algorithms to well-known one-class methods, and the experimental analysis are conducted on real datasets.