Is there a minimum data set for making classification decisions with Naïve Bayes? I have objects from two class  -- 11 objects in each class. Where each object has 3 features. with such small dataset, does it make any sense to train & use Naïve Bayes to classify new objects (test cases) ? 
 A: I believe I have found the answer for this one: it was shown in 2001 by A. Ng: "However, it was not too long ago that many people preferred logistic regression(a discriminative model) to naive Bayes(a generative model). In fact, many people preferred discriminative to generative models in general, since discriminative models achieve lower asymptotic errors. These preferences were tested by Ng and Jordan, and it was shown that a decision between which classifier is better is not always a simple answer; it was shown that naive Bayes reaches its asmyptotic error very quickly with regards to the number of training examples."
A: Support vector machines typically outperform other classification methods when training data is sparse (This is based on my own experiences as well as those of my colleagues).
I haven't seen much published work analyzing the question of which classifier is best besides this paper in ICML 2006: "An empirical comparison of supervised learning algorithms". Unfortunately, they keep the  # of training samples constant so this will not answer your question directly, but it may provide some good references if you want to read more in depth on the subject.
