Discriminative Learning of Bayesian Network Classifiers via the TM Algorithm

The learning of probabilistic classification models can be approached from either a generative or a discriminative point of view. Generative methods attempt to maximize the unconditional log-likelihood, while the aim of discriminative methods is to maximize the conditional log-likelihood. In the case of Bayesian network classifiers, the parameters of the model are usually learned by generative methods rather than discriminative ones. However, some numerical approaches to the discriminative learning of Bayesian network classifiers have recently appeared. This paper presents a new statistical approach to the discriminative learning of these classifiers by means of an adaptation of the TM algorithm [1]. In addition, we test the TM algorithm with different Bayesian classification models, providing empirical evidence of the performance of this method.