Interest in the naive Bayes learning algorithm within
machine learning circles can be attributed to Clark
and Niblett's paper on the CN2 rule learner ( Clark
& Niblett, 1989). In this paper they included a simple
Bayesian classifier ( naive Bayes) as a "straw man"
in their experimental evaluation and noted its good
performance compared to more sophisticated learners.
Although it has been explained how naive Bayes can
work well in some cases where the attribute independence
assumption is violated ( Domingos & Pazzani,
1997) the fact remains that probability estimation is
less accurate and performance degrades when attribute
independence does not hold.