Active Learning by the Naive Credal Classifier
Authors: Alessandro Antonucci, Giorgio Corani and Sandra Gabaglio
Abstract: In standard classification a training set of supervised instances is given. In a more general setup, some supervised instances are available, while further ones should be chosen from an unsupervised set and then annotated. As the annotation step is costly, active learning algorithms are used to select which instances to annotate to maximally increase the classification performance while annotating only a limited number of them. Several active learning algorithms are based on the naive Bayes classifier. We work instead with the naive credal classifier, namely an extension of naive Bayes to imprecise probability. We propose two novel methods for active learning based on the naive credal classifier. Empirical comparisons show performance comparable or slightly superior to that of approaches solely based on the naive Bayes.
Details: In Cano, A. and Gómez-Olmedo, M. and Nielsen, T.D. (Eds.), Proceeding of the sixth European Workshop on Probabilistic Graphical Models (PGM 2012).
A version similar to the published paper can be downloaded.