Re: sparse bigrams for hyperspace

      how many errors must usually be corrected before crm reaches 80%, 90%,
      95% and 99% accuracy?

      are we talking hundreds, thousands or tens of thousands?

    Hundreds, or less, for 90%.  The SpamAssassin corpus trains about 500
    documents, and we get about 98 to 99% accuracy on the "tenfold
    validate torture test" on most of the classifiers just with judicious
    choices of learning method and threshold thickness.


thanks!  i have trained my newspaper article classifier (9 categories
+ 1 fail-category) on about 2300 articles classified by hand, and my
accuracy is never above 53% with hyperspace. osbf fails to yield more
than 28% accuracy.



    Past that, I have less than 2000 documents *total* in
    my Reavercache of known examples, and probably have purged away
    twice that many, and am pushing 99.99% accuracy on my real email.


      does the hyperspace take the order of features in a text into account,
      and would it be useful to arrange the features in a text into
      five-grams, or perhaps sparse bigrams, like osbf does?

    Somewhat.  Default hyperspase uses sparse bigrams, like OSB and
    OSBF do.  Beyond that, it doesn't care about order of features
    (in fact, to optimize classification, it hashes them and then
    sorts the hashes to get O(n) classification time).


again, thanks! i was hoping to improve accuracy by using sparse
bigrams. so much for that :)

i might have better luck using some meta-tagged material we have lying
around, and putting the wordclass, along with the lemma into the
classifier as a single two-word feature.

------------------------------------------------------------------------------
Come build with us! The BlackBerry® Developer Conference in SF, CA
is the only developer event you need to attend this year. Jumpstart your
developing skills, take BlackBerry mobile applications to market and stay 
ahead of the curve. Join us from November 9-12, 2009. Register now!
http://p.sf.net/sfu/devconf