Irrelevance and ParameterLearning in Bayesian Networks

Nevin Lianwen Zhang

Abstract: It is possible to learn the parameters of a given Bayesian network structure from data because those parameters influence the probability of observing the data. However, some of the parameters are irrelevant to the probability of observing a particular data case. This paper shows how such irrelevancies can be exploited to speedup various algorithms for parameter learning in Bayesian networks. Experimental results with one of the algorithms, namely the EM algorithm, are presented to demonstrate the gains of this exercise.