Pruning Ensembles of One-Class Classifiers with X-means Clustering

Abstract

In this paper, we present a novel approach for pruning ensembles of one-class classifiers. One-class classification is among the most challenging topics in the contemporary machine learning. Creating multiple classifier systems for this task is one of the most effective ways of improving the quality and robustness in case of lack of counterexamples. However, very often we are faced with the problem of redundant or weak classifiers in the pool, as one-class ensembles tend to overproduce the base learners. To tackle this problem a dedicated pruning scheme must be employed, which will allow to discard classifiers that do not contribute to the formed ensemble. We propose to approach this problem as a clustering task. We discover groups of classifiers according to their support function values for the target class. For each group, we select the most representative classifier and discard the remaining ones. We apply an efficient x-means clustering algorithm, that automatically establishes the optimal number of clusters with the use of the Bayesian Information Criterion. Experimental results carried out on a set of benchmarks prove, that our proposed method is able to provide an efficient pruning mechanism for one-class problems.