Influence of Distance Measures on the Effectiveness of One-Class Classification Ensembles
Bartosz Krawczyk, Michał Woźniak · Applied Artificial Intelligence · 2014
Because of the presence of only a single class during the learning procedure, one-class classification is considered as one of the most challenging tasks in contemporary machine learning. Recently, ensemble approaches have been applied to this task, proving that they may lead to a significant improvement of the robustness of the recognition system. Most of 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, one must map the distance into support function (probability). Additionally, we use weighted one-class support vector machine, which utilizes a distance-based function for calculating weights assigned to objects. Therefore, one may easily see that the measure used has a crucial impact on the quality of base classifiers as well as on the classifier fusion. This study investigates the performance of seven different distance measures over a set of diverse benchmark datasets. Additionally, we analyze the correlation between the used distance measure and the selected fusion method. Experimental analysis allows us to shed some light on strengths and weaknesses of examined distance metrics applied to combining one-class classifiers.