An Efficient Global Discretization Method Technical Report Number 296

The development of an effective and efficient method for discretization of continuous variables is an important problem to be solved in developing generally applicable methods for data mining. In Ho and Scott 1997, we describe a new technique for discretization of continuous variables based on zeta, a measure of strength of association between nominal variables. The old zeta method partitions a continuous variable into as many sub-ranges as the values of the classification variable. In this paper, we introduce a generalised form of zeta that remove this restriction. The new method is also more efficient particularly when there are more than 4 distinct values in the classification variable. A series of experimental evaluations were performed comparing the improved zeta, together with the old zeta and C4.5 which employs its own local discretization method. Surprisingly, there were little difference in accuarcies produced by all the 3 methods, however, the improved zeta discretization technique runs considerably faster in all cases. We conclude that zeta global discretization offers advantages over C4.5.