Growing decision trees in an ordinal setting.

Although ranking (ordinal classification/regression) based on criteria is related closely to classification based on attributes, the development of methods for learning a ranking on the basis of data is lagging far behind that for learning a classification. Most of the work being done focuses on maintaining monotonicity (sometimes even only on the training set). We argue that in doing so, an essential aspect is mostly disregarded, namely, the importance of the role of the decision maker who decides about the acceptability of the generated rule base. Certainly, in ranking problems, there are more factors besides accuracy that play an important role. In this article, we turn to the field of multicriteria decision aid (MCDA) in order to cope with the aforementioned problems. We show that by a proper definition of the notion of partial dominance, it is possible to avoid the counter-intuitive outcomes of classification algorithms when applied to ranking problems. We focus on tree-based approaches and explain how the tree expansion can be guided by the principle of partial dominance preservation, and how the resulting rule base can be graphically represented and further refined. © 2003 Wiley Periodicals, Inc. [ABSTRACT FROM AUTHOR]