Combining One-Class Classifiers via Meta-Learning

Eitan Menahem Lior Rokach, Yuval Elovici

arXiv preprint arXiv:1112.5246, 2011

We examine various methods for combining the output of one-class models. In particular, we show that simple meta-learning based ensemble achieves better result than weighting methods. Furthermore we propose a new one-class ensemble scheme, called TUPSO that uses metalearning for combining multiple one-class classifiers. We also present a new one-class classification performance measures to weigh the base-classifiers, a process that proved helpful for increasing the classification performance of the induced ensemble. Our experimental study shows that the proposed method significantly outperforms exiting methods.