AN EXTENSION OF THE Q DIVERSITY METRIC FOR INFORMATION PROCESSING IN MULTIPLE CLASSIFIER SYSTEMS: A FIELD EVALUATION
Filippo Sciarrone · International Journal of Wavelets Multiresolution and Information Processing · 2013
Nowadays, in pattern recognition and classification, does not exist a dominant classifier for all data distributions. Also, data distribution of the task at hand is usually unknown, i.e. there is no algorithm achieving the best accuracy for all situations. An answer and a challenge to this problem is to build ensembles of classifiers working together, i.e. Multiple Classifier Systems, instead of building and running different classifiers separately: Multiple Classifier Systems can show better performance than a single classifier, provided a careful choice of the individual classifiers composing it. Furthermore, diversity among single classifiers, measured by some diversity metrics, is known to be a necessary condition to improve the ensemble performance. In this paper we extend the use of one of the most used diversity metrics, that is the Q diversity metric, from an oracle output to a soft output for the choice of the best classifier ensemble. We introduce the Qt diversity metric, i.e. an extension of the Q metric to multi-label and multi-ranking Multiple Classifier Systems. A field evaluation of this metric is presented in a text categorization case study, using as a test set the standard document corpus Reuters 21578 ModApte 10. Our results strengthen the use of the extended metric in multi-label and multi-ranking Multiple Classifier Systems.