On using Bayesian networks for complexity reduction in decision trees