Discretization Of Continuous-valuedFeatureVariablesInNaiveBayesianNetworks
Room: BBL 065
Speaker: Roel Bertens
Title: Discretization of Continuous-valued Feature Variables in Naive Bayesian Networks
In Bayesian networks all variables should be discrete. Therefore all continuous-valued domain variables should be discretized when building a Bayesian network. To this end, several methods with different input parameters exist, which can all have a considerable impact on the posterior distributions computed from a network.
The consequences for the posterior distribution of the class variable in a naive Bayesian network, of changing the discretization for a feature variable, are presented in this talk. Using the concepts from sensitivity analyses we will determine some conditions that need to hold in order to change the behavior of a naive Bayesian classifier as a result of a different discretization for a feature variable. These conditions comprise bounds for the values of the prior class variable distribution and they constrain the relationship between the parameters from the conditional probability table of the variable that is to be discretized.