A novel approach to fully representing the diversity in conditional dependencies for learning Bayesian network classifier