Like most other generative models, Bayesian networks are commonly learned using generative approaches. In particular, the maximum likelihood approach is often used to produce structures and to estimate the parameters of Bayesian nets. Besides their important application in modeling, Bayesian nets also always had important applications on specialized tasks such as diagnosis or classification. It is therefore reasonable to use discriminative approaches that directly maximize Bayesian nets' performance on these tasks.; In this dissertation, we describe new discriminative approaches, with a focus on maximum conditional likelihood estimation. We compare this approach with the commonly used maximum likelihood approach. We provide empirical evidence to show that the discriminative approaches indeed perform better than the generative approach. In addition, the empirical results show that the discriminative approaches perform well on many real world problems.
展开▼
