Fix for #4 - no prior used in Bayes Classifier

[bmuller]

Without taking the prior probabilities for each class under consideration, you have a classifier that uses only likelihood (more specifically, log likelihood) rather than a classifier that uses Bayes theorem.

[dabble]

Thank you.

I spent a while reviewing http://en.wikipedia.org/wiki/Naive_Bayes_classifier
to better understand this change.

My feedback would be to update your change to replace spaces with tabs as
the original file is mostly tabbed. [Note: I don't have rights to this repository.]

The only downside I see to accepting this change is if someone called train_
multiple times per document instead of 1 time per document.

[bmuller]

Calling train more than once per document would result in bias even for a classifier just using log likelihoods.

Right now, the classifier uses only the likelihood to estimate the posterior - which means it's not a naive bayes classifier at all. The prior must be taken into account.

Since it looks like this rather egregious bug is not going to be fixed - I suggest using https://github.com/livingsocial/ankusa

[dabble]

Actually, I was thinking of calling train_ once each for title, author, body of a document.
Since without this fix, it only counted words, this would not affect the outcome.

The only reason I had for doing it (and I didn't) was to avoid concatenating strings only
to break them apart again.

As you point out, without this, this isn't really naive bayes. (I'm not sure counting words
multiple times per document is really naive bayes either...)

I hadn't run across the ankusa classifier. Thanks.

[cardmagic]

I am sorry it took so long to merge this in