Export Reviews, Discussions, Author Feedback and Meta-Reviews

This paper focusses on developing an efficient transductive ensemble learning algorithm that follows the theoretical work of [1]. The problem, that is characterized as a zero-sum game between a predictor and nature, is first reviewed. The authors then introduce a generalization to the scenario where the voters of the ensemble are "specialists": voters that can abstain on a subset of the examples. An instanciation of the method to specialists learned using a Random Forest (named HedgeClipper) is devised from the theoretical framework, and empirically evaluated on real world datasets. This paper is very well written and its original contributions are made very clear by the authors.