Development and Design of General Data Mining System

摘要

In this paper, we focus on top-down discretization methods and propose a new method for supervised discretization based on class-feature correlation by defining a class-feature contingency factor. The proposed method takes into consideration the distribution of all samples to generate an ideal discretization scheme. The method maintains a high interdependence between the target class and the discretized attribute, and avoids overfitting. Empirical evaluation of seven discretization algorithms on UCI real datasets show that the novel algorithm can yield a better discretization scheme that improves the accuracy of decision tree classification. As to the execution time of discretization and the number of generated rules, our approach also achieves promising results.