Fuzzy Discretization for Data Mining

Data preprocessing or data preparation is the most time consuming and most laborious step in the whole data mining process. The reason for data preprocessing is twofold; it is necessary to select (or create) from available data characteristics relevant for given data mining task, and to represent these characteristics in a form suitable for selected data mining algorithm. Among the typical data operations performed in this step discretization of numeric attributes plays an important role as algorithms for creating either association or decision rules cannot handle numeric attributes directly. Different approaches to discretization can be used. Equidistant or equifrequent discretization are typical examples of so-called "class-blind" methods since they deal only with the discretized attribute. Another group of discretization are "class-sensitive" methods; the fact that the examples (objects) belong to different classes is taken into account here. The discretization procedures typically generate sharp boundaries (thresholds) between intervals. The paper describes a class-sensitive discretization method where the boundary between intervals is defined using fuzzy membership function. The paper shows an experimental evaluation of the proposed method on some benchmark data and also compares the proposed method with more standard (i.e. crisp) discretization. A naive Bayesian classifier as well as a tree learning algorithm are used in the experiments.