Experiments on distance measures for combining one-class classifiers

The paper investigates the influence of different types of distance measures on the performance of a multiple classifier system consisting of one-class classifiers. This specific type of machine learning approach uses examples only from a single class to derive a decision boundary - hence its is often referred to as learning in the absence of counterexamples. Combining several one-class classifiers is a promising research direction, as it often results in a more precise classification than when using just a single model. Most one-class classifiers base their decision on a distance from an object to the decision boundary, canonically expressed in the Euclidean measure. When combining such predictors it is necessary to map the distance into probability, therefore the measure used has a crucial impact on the classifier fusion. This paper proposes alternative distance measures for one-class classification, which are evaluated through experimental investigations.