Re: [Classifier4j-devel] c4J vs. PopFile

[moedusa <mo...@in...>]

Matt Collier wrote:

> I think I just figured out the primary difference between what POPFile is 
> doing and what we are currently doing.  
> 
> POPfile is keeping track of how many messages have been trained in each 
> category as well as an overall message count.  I believe it is this additional 
> information that is allowing them to calculate additional probabilities for 
> muti-category sorting.
> 
> Does this sound reasonable?

I think, yes. Though, I have no idea, how, even with this kind of 
metainformation, it could produce preciese results... But I know, that 
any metainformation is very useful :)

Also it is interesting for me, is it possible somehow to notify 
something when classifier is not sure what category to assign. Let's say 
when probability is neutral, could'nt it rise an event that the text was 
missed, so operator could train it. It seems useful for me, because if 
you use classifier to classify, it must classify, and if it does not 
know what to do, it should just ask... Well, I suggest to think about 
how to make it a little bit more self-learning or something...

Of course, it is possible to  catch neutral probability from external 
code, but I think, that it would be a nice option in API also. Just a 
thought, nothing more.