Learning Parameters in Discrete Naive Bayes Models by Computing Fibers of the Parametrization map

Discrete Naive Bayes models are usually defined parametrically with a map from a parameter space to a probability distribution space. First, we present two families of algorithms that compute the set of parameters mapped to a given discrete Naive Bayes distribution satisfying certain technical assumptions. Using these results, we then present two families of parameter learning algorithms that operate by projecting the distribution of observed relative frequencies in a dataset onto the discrete Naive Bayes model considered. They have nice convergence properties, but their computational complexity grows very quickly with the number of hidden classes of the model.