Abstract
In recent years, the introduction of aggregation methods led to many new techniques within the field of prediction and classification. The most important developments, bagging and boosting, have been extensively analyzed for two- and multiclass problems. While the proposed methods treat the class indicator as a nominal response without any structure, in many applications the class may be considered as an ordered categorical variable. In this article, variants of bagging and boosting are proposed, which make use of the ordinal structure. It is demonstrated how the predictive power is improved by the use of appropriate aggregation methods. Comparisons between the methods are based on misclassification rates as well as criteria that take ordinality into account, like absolute or squared distance measures.