Dynamic Discretization: A Combination Approach

Supervised discretization refers to the problem of transforming continuous attributes of a decision table into discredited ones. It is important for some artificial intelligence theories where nominal data are required or preferred. Instead of depending on the experience of human experts, supervised discretization algorithms learn from the data. However, the results of such algorithms may be sensitive to the change of the data. In this paper, we propose to compute more stable and informative discretization schemes through subtable sampling and scheme combination. Discretization schemes computed in this way are called dynamic discretization schemes. Experimental results on some well-known datasets show that they are helpful for obtaining decision rules with better accuracy and F-measure.