Adapting Decision DAGs for Multipartite Ranking

Abstract

Multipartite ranking is a special kind of ranking for problems in which classes exhibit an order. Many applications require its use, for instance, granting loans in a bank, reviewing papers in a conference or just grading exercises in an education environment. Several methods have been proposed for this purpose. The simplest ones resort to regression schemes with a pre- and post-process of the classes, what makes them barely useful. Other alternatives make use of class order information or they perform a pairwise classi cation together with an aggregation function. In this paper we present and discuss two methods based on building a Decision Directed Acyclic Graph (DDAG). Their performance is evaluated over a set of ordinal benchmark data sets according to the C-Index measure. Both yield competitive results with regard to stateof- the-art methods, specially the one based on a probabilistic approach, called PR-DDAG