Pruning One-Class Classifier Ensembles by Combining Sphere Intersection and Consistency Measures

Abstract

One-class classification is considered as one of the most challenging topics in the contemporary machine learning. Creating Multiple Classifier Systems for this task has proven itself as a promising research direction. Here arises a problem on how to select valuable members to the committee - so far a largely unexplored area in one-class classification. This paper introduces a novel approach that allows to choose appropriate models to the committee in such a way that assures both high quality of individual classifiers and a high diversity among the pool members. We aim at preventing the selection of both too weak or too similar models. This is achieved with the usage of an multi-objective optimization that allows to consider several criteria when searching for a good subset of classifiers. A memetic algorithm is applied due to its efficiency and less random behavior than traditional genetic algorithm. As one-class classification differs from traditional multi-class problems we propose to use two measures suitable for this problem - consistency measure that allow to rank the quality of one-class models and introduced by us sphere intersection measure that serves as a diversity metric. Experimental results carried on a number of benchmark datasets proves that it outperforms traditional single-objective approaches.