Discriminative learning of the parameters in the naive Bayes model is known to be equivalent to a logistic regression problem. Here we show that the same fact holds for much more general Bayesian network models, as long as the corresponding network structure satisfies a certain graph-theoretic property. The property holds for naive Bayes but also for more complex structures such as tree-augmented naive Bayes (TAN) as well as for mixed diagnostic-discriminative structures. Our results imply that for networks satisfying our property, the conditional likelihood cannot have local maxima so that the global maximum can be found by simple local optimization methods. We also show that if this property does not hold, then in general the conditional likelihood can have local, non-global maxima. We illustrate our theoretical results by empirical experiments with local optimization in a conditional naive Bayes model. Furthermore, we provide a heuristic strategy for pruning the number of parameters and relevant features in such models. For many data sets, we obtain good results with heavily pruned submodels containing many fewer parameters than the original naive Bayes model.

Classification problems have a long history in the machine learning literature. One of the simplest, and yet most consistently well performing set of classifiers is the Nave Bayes models. However, an inherent problem with these classifiers is the assumption that all attributes used to describe an instance are conditionally independent given the class of that instance. When this assumption is violated (which is often the case in practice) it can reduce classification accuracy due to "information double-counting" and interaction omission.

One of the simplest, and yet most consistently well-performing set of classifiers is the naïve Bayes models (a special class of Bayesian network models). However, these models rely on the (naïve) assumption that all the attributes used to describe an instance are conditionally independent given the class of that instance. To relax this independence assumption, we have in previous work proposed a family of models, called latent classification models (LCMs). LCMs are defined for continuous domains and generalize the naïve Bayes model by using latent variables to model class-conditional dependencies between the attributes. In addition to providing good classification accuracy, the LCM model has several appealing properties, including a relatively small parameter space making it less susceptible to over-fitting. In this paper we take a first-step towards generalizing LCMs to hybrid domains, by proposing an LCM model for domains with binary attributes. We present algorithms for learning the proposed model, and we describe a variational approximation-based inference procedure. Finally, we empirically compare the accuracy of the proposed model to the accuracy of other classifiers for a number of different domains, including the problem of recognizing symbols in black and white images. 1