There is one caveat. In order to avoid the zerofrequency
problem our implementation of naive Bayes
uses the Laplace estimator to estimate the conditional
probabilities for nominal attributes and this interacts
with the weighting scheme. We found empirically that
it is opportune to scale the weights so that the total
weight of the instances used to generate the naive
Bayes model is approximately k. Assume that there
are r training instances x; with d; :