Incremental Augmented Naive Bayes Classifiers

We propose two general heuristics to transform a batch Hill-climbing search into an incremental one. Our heuristics, when new data are available, study the search path to determine whether it is worth revising the current structure and if it is, they state which part of the structure must be revised. Then, we apply our heuristics to two Bayesian network structure learning algorithms in order to obtain incremental Augmented Naive Bayes classifiers. We experimentally show that our incremental approach saves a significant amount of computing time while it yields classifiers of similar quality as the ones learned with the batch approach.