Aprendizaje discriminativo de clasificadores Bayesianos

The learning of Bayesian network models for classification is usually approached from a generative point of view. That is, the learning process attempts to maximize the likelihood of the dataset given the learned model. However, there is another approach, the discriminative learning, which attempts to maximize the conditional likelihood. This discriminative learning seems to be a more natural approach for classification purposes. Nevertheless, generative approaches can sometimes yield better results than discriminative ones. Some methods for the discriminative learning of Bayesian network classifiers have recently appeared in the literature. In this paper, we present a new method for the discriminative learning of both structure and parameters for Bayesian network classifiers based on an adaptation of the TM algorithm. Additionally, we present an empirical evaluation of the method proposed using different datasets from UCI repository