Meta-prediction of semi-naive Bayesian network classifiers based on dataset complexity characterization
M. Julia Flores, José Antonio Gámez · 2012
Ever since naive Bayes was proposed, many have been the attempts to try to alleviate its naive assumption to obtain better accuracy records, without further increasing its complexity. In this line we can find a group of Bayesian network classifiers that either do not perform structural search or it is very simple, known as the family of semi-naive Bayesian network classifiers (BNCs). Given a particular dataset, it would be desirable, based on the characteristics presented, to find out which semi-naive BNCs obtains the best possible expected prediction, since estimations based solely on expected accuracy on training can be misleading. In this paper we propose an automatic procedure to carry out this meta-prediction process, based on the values of several data complexity measures for supervised classification. We resort to multi-label classification to test this procedure, obtaining promising results.