Simultaneous Partitioning of Input and Class Variables for Supervised Classification Problems with Many Classes

Abstract In the data preparation phase of data mining, supervised discretization and value grouping methods have numerous applications: interpretation, conditional density estimation, filter selection of input variables, variable recoding for classifi-cation methods. These methods usually assume a small number of classes, typically less than ten, and reach their limit in case of too many classes. In this paper, we extend discretization and value grouping methods, based on the partitioning of both the input and class variables. The best joint partitioning is searched by maximiz-ing a Bayesian model selection criterion. We show how to exploit this preprocess-ing method as a preparation for the naive Bayes classifier. Extensive experiments demonstrate the benefits of the approach in the case of hundreds of classes.