Instance Reduction for One-Class Classification

Abstract

Instance reduction techniques are data preprocessing methods originally developed to enhance the nearest neighbor rule for standard classification. They reduce the training data by selecting or generating representative examples of a given problem. These algorithms have been designed and widely analyzed in multi-class problems providing very competitive results. However, this issue was rarely addressed in the context of one-class classification. In this specific domain a reduction of the training set may not only decrease the classification time and classifier's complexity, but also allows us to handle internal noisy data and simplify the data description boundary. We propose two methods for achieving this goal. The first one is a flexible framework that adjusts any instance reduction method to one-class scenario by introduction of meaningful artificial outliers. The second one is a novel modification of evolutionary instance reduction technique that is based on differential evolution and uses consistency measure for model evaluation in filter or wrapper modes. It is a powerful native one-class solution that does not require an access to counterexamples. Both of the proposed algorithms can be applied to any type of one-class classifier. On the basis of extensive computational experiments, we show that the proposed methods are highly efficient techniques to reduce the complexity and improve the classification performance in one-class scenarios.