[spambayes-dev] ReleBayes - Food For Tought?

[Janne Sinkkonen]

On Tuesday 18 May 2004 08:34, Eicker | eWerx!..communications° wrote:

> What's on my wish list now is "ReleBayes": I mean a filter that not
> only kills spam (that's SpamBayes) but a filter that sorts ham-mails
> by relevancy. It would be even easier for the user to use and even
> more powerful:
> 
> - Relevancy should be learned when a user *replies* to an email. I
> believe you respond to 15-25% of your eMails only like I do.

I have done this, running two incarnations of spambayes (in Linux). I split 
mail to two folders, one being for probably non-replied mail and one for 
replied. This kind of works, at least well enough that I want to stick with 
it. Implementation is simply enough - just run two spambayes with different 
initialization files, and tune options to give different header names etc.

I agree that there is potential in automatic text analysis and classification. 
Spambayes would be a good framework for trying various kind of 
new probabilistic text analysis techniques (Latent Dirichlet Allocation, 
multinomial PCA, etc.) with some kind of discriminative approach.

-- 
Janne