This paper presents a lazy approach to learning naive
Bayes models. Like all lazy learning methods our
approach simply stores the training data and defers
the effort involved in learning until classification time.
When called upon to classify a new instance, we construct
a new naive Bayes model using a weighted set of
training instances in the locale of the test instance. Local
learning helps to mitigate the effects of attribute
dependencies that may exist in the data as a whole
and we expect this method to do well if there are no
strong dependencies within the neighbourhood of the
test instance. Because naive Bayes requires relatively
little data for training, the neighbourhood can be kept
small, thereby reducing the chance of encountering
strong dependencies. In our implementation the size of
the neighbourhood is chosen in a data-dependent fashion
based on the distance of the k-th nearest-neighbour