DbEr: A New Discretization Algorithm Based on the Entropy of Relation

Abstract

We propose a new discretization method DbEr, which is founded on the entropy of relation(ER). DbEr includes a bottom-up and a top-down discretization methods at the same time. Compared with other methods of discretization, DbEr takes ER as the only measure for not only the choice criterions of attributes and cutpoints, but also the stopping criterion of discretization. Using a single measure as the criterions in the whole process of discretization will maintain the consistency of preference. This is an important merit of DbEr. Furthermore, some new concepts are put forward to discuss the advantage of taking ER as the measure for discretization. It is shown that the maximal value of ER is the compromise of unknown knowledge and inconsistent knowledge. At last, we use naive Bayesian classifier as classification tools to compare DbEr with some other discretization algorithms. The result shows that DbEr can do better than most of other discretization algorithms.