On Tue, Mar 25, 2008 at 3:49 PM, Isabel Drost <[EMAIL PROTECTED]> wrote:
> The paper looks interesting: The modifications to naive bayes presented in > the > paper seem to lead to a classifier that is comparable to SVM performance > for > text classification while having far better performance. The text transforms aren't particularly interesting---they're essentially just a TFIDF pre-processing step. This botches the multinomial model a bit, but works well and has some theoretical motivation (in sec. 4). I thought the discovery of the "skewed data bias" was quite interesting. Was never completely satisfied with our solution, Complement Naive Bayes. But, it did seem to solve the problem and yield improved performance for data where the number of examples per class varied widely. Seems like you need an algorithm that outputs comparable scores for each > document and is neither under- nor overconfident. I remember vaguely that > the > vanilla NB had some problems in this respect. Complement NB gets rid of some of this problem, though Logistic Regression or Softmax (the multiclass variant) is probably a generally better solution. 'course LR/Softmax requires optimization whereas (C)NB requires little more than counting and basic math ops... easier to implement... Jason -- Jason Rennie Head of Machine Learning Technologies, StyleFeeder http://www.stylefeeder.com/ Samantha's blog & pictures: http://samanthalyrarennie.blogspot.com/
