On diversity measures for fuzzy one-class classifier ensembles
Bartosz Krawczyk, Michał Woźniak · 2013
One-class classification became one of the most challenging research areas of the contemporary machine learning. Contrary to canonical task here we have only information about a single class at our disposal. Therefore more sophisticated methodologies, that are able to handle all the nuisances of the target distribution are required. Fuzzy logic seems an attractive solution to handle imprecision and to naturally grade the influence of the input data on the decision boundary. In this paper we propose to create a committee of fuzzy one-class support vector machines based on the random subspace method and diversity-based ensemble pruning technique. We investigate if there is a difference when using crisp or fuzzy diversity measures - and if so, then which of them are preferable for fuzzy one-class ensembles. The experimental investigations carried on a wide selection of benchmark datasets and backed-up with a statistical test of significance proves that selecting a proper diversity measure for fuzzy one-class ensemble has a crucial impact on its overall quality.