Discretization

Discretization consists in replacing an originally continuous attribute by a discrete attribute, with different values assigned to particular intervals of the original attribute's range. This chapter outlines the motivation behind discretization and discusses the definition of the task in more detail. The ultimate verification of the effects of discretization is always provided by the evaluation of models obtained when performing the target task using discretized data. Unsupervised discretization algorithms, which do not take any target attribute into account, are based entirely on the observed distribution of the continuous attribute to be discretized in the training set. Supervised discretization requires considerably more effort than the simple unsupervised algorithms. Since discretization is a part of the modeling process, its effects have to be therefore evaluated using an independent data subset. This can be accomplished by employing a selected evaluation procedure.