The classification of patterns into naturally orderedlabels is referred to as ordinal regression. This paper proposesan ensemble methodology specifically adapted to this type ofproblems, which is based on computing different classificationtasks through the formulation of different order hypotheses.Every single model is trained in order to distinguish betweenone given class (k) and all the remaining ones, but groupingthem in those classes with a rank lower than k, and thosewith a rank higher than k. Therefore, it can be considered asa reformulation of the well-known one-versus-all scheme. Thebase algorithm for the ensemble could be any threshold (oreven probabilistic) method, such as the ones selected in thispaper: kernel discriminant analysis, support vector machinesand logistic regression (all reformulated to deal with ordinalregression problems). The method is seen to be competitive whencompared with other state-of-the-art methodologies (both ordinaland nominal), by using six measures and a total of fifteen ordinaldatasets. Furthermore, an additional set of experiments is used tostudy the potential scalability and interpretability of the proposedmethod when using logistic regression as base methodology forthe ensemble.
展开▼