Hi,

I have unusual idea for different evaluation of scores or exactly another 
method of mail SPAMness rating.

Idea is based on using Bayesian theorem. Contrary to usual two-level score 
computation, where first level is Bayesian classification of word tokens from 
mail contents together with evaluation of other rules and second level is 
weighted combining (sum) of results, my idea consists on including matching 
rules as additional tokens into Bayesian evaluation of that mail. Decision if 
mail is or not SPAM will be made by testing of hypothesis "is SPAM" on selected 
level (e.g. 95%). Naturally this method is functional only when Bayes database 
is initially prefilled with some reasonable values for rule-tokens which is 
analogy to rule scores.

I hope, this method have advantages in
- adaptivity when learning is used
- simplicity because of only single evaluating method
- it is founded on regular statistical test not simplistic comparison of sum of 
weights

Try to weight up my idea please.

Thanks for your comments.

M. Vancl

Reply via email to