RE: [Classifier4j-devel] surprising classify score 0.01

[Nick L. <nl...@es...>]

> 
> I am enjoying experimenting with your Classifer4J 0.5,
> but I ran across a result I did not expect.  I have
> trained a BayesianClassifier with 22 positive examples
> and 1600 negative examples.  Many of the positive
> examples contain the word "http".  None of the
> negative examples contain this word.  
> 
> The surprising result is that the score of a sentence
> with "http" is 0.01.  Can you help me to understand
> why?
> 
> Here is the sentence and the WordProbability
> probabilities for each of the words in the sentence
> that were in the training data:
> 
> score = 0.01 for "Mozilla/4.0 (compatible;
> grub-client-1.3.7; Crawl your own stuff with
> http://grub.org)"
>  0.11822660098522167 Mozilla
>  0.020618556701030927 4
>  0.07223476297968397 0
>  0.029239766081871343 compatible
>  0.10619469026548672 1
>  0.5454545454545454 3
>  0.01 7
>  0.99 http
> 

That sound right to me

The math goes like this (and I'm going to round those number off, because I
can't be bothered typing them into my calculator):

score =
((0.11)(0.02)(0.07)(0.02)(0.11)(0.55)(0.01)(0.99))/((0.11)(0.02)(0.07)(0.02)
(0.11)(0.55)(0.01)(0.99) + (1 - 0.11)(1 - 0.02)(1 - 0.07)(1 - 0.02)(1 -
0.11)(1 - 0.55)(1 - 0.01)(1 - 0.99))


= 0.000000001844766/(0.000000001844766 +
(0.89)(0.98)(0.93)(0.98)(0.89)(0.45)(0.99)(0.01))

= 0.000000001844766/(0.000000001844766 + 0.003151830266046)

= 0.000000001844766/0.003151832110812

= 0.00000058529957660871

Classifier4J has a cut-off system where anything under 0.01 gets 0.01.

Does that help explain things?

This code for this is in net.sf.classifier4J.bayesian.BayesianClassifier
(see
<http://classifier4j.sourceforge.net/xref/net/sf/classifier4J/bayesian/Bayes
ianClassifier.html>)


Nick Lothian