Uncertainty sampling methods for one-class classifiers

Selective sampling, a part of the active learning method, reduces the cost of labeling supplementary training data by asking for the labels only of the most informative, unlabeled examples. This additional information added to an initial, randomly chosen training set is expected to improve the generalization performance of a learning machine. We investigate some methods for a selection of the most informative examples in the context of one-class classification problems (OCC) i.e. problems where only (or nearly only) the examples of the so-called target class are available. We applied selective sampling algorithms to a variety of domains, including realworld problems: mine detection and texture segmentation. The goal of this paper is to show why the best or most often used selective sampling methods for two- or multi-class problems are not necessarily the best ones for the one-class classification problem. By modifying the sampling methods, we present a way of selecting a small subset from the unlabeled data to be presented to an expert for labeling such that the performance of the retrained one-class classifier is significantly improved.