Mapping regression into classification was first proposed by Weiss & Indurkhya (1993, 1995). These authors incorporate the mapping within their regression system. They use an algorithm called P-class that splits the continuous values into a set of K intervals, and use cross validation to estimate the number of intervals. Their methodology is similar to our KM+VNI discretization. Compared to this work we added other alternative discretization methods and empirically proved the necessity of a search-based approach to class discretization. Moreover, by separating the discretization process from the learning algorithm we extend this approach to other systems. Finally, we have introduced the use of misclassification costs to overcome the inadequacy of classification systems to deal with ordinal target variables.
Previous work on continuous attribute discretization usually proceeds by trying to maximize the mutual information between the resulting discrete attribute and the classes (Fayyad & Irani, 1993). This strategy is applicable only when the classes are given. Ours is a different problem, as we are determining which classes to consider.