Entropy and MDL discretization of continuous variables for Bayesian belief networks

An efficient algorithm for partitioning the range of a continuous variable to a discrete number of intervals, for use in the construction of Bayesian belief networks (BBNs), is presented here. The partitioning minimizes the information loss, relative to the number of intervals used to represent the variable. Partitioning can be done prior to BBN construction or extended for repartitioning during construction. Prior partitioning allows either Bayesian or minimum descriptive length (MDL) metrics to be used to guide BBN construction. Dynamic repartitioning, during BBN construction, is done with a MDL metric to guide construction. The methods are demonstrated with data from two epidemiological studies and these results are compared for all of the methods. The use of the partitioning algorithm resulted in more sparsely connected BBNs, than with binary partitioning, with little information loss from mapping continuous variables into discrete ones. ©2000 John Wiley & Sons, Inc.