Thread: [Classifier4j-devel] Bayesian Results 0.99 for Everything?

Status: Beta

Brought to you by: nicklothian

classifier4j-devel

[Classifier4j-devel] Bayesian Results 0.99 for Everything?

From: Brent L J. <br...@bj...> - 2004-01-02 22:14:39

How does the BayesianClassifier differ from a program similar
to "SpamBayes" (has an Outlook plugin that uses the Bayesian
algorith to classify emails).  I "taught" the Bayesian classifier by
teaching
it with some spam emails I have.  Now it returns a 0.99 classify
result for practically EVERYTHING.

A little background..

I exported about 6400 spam emails from Outlook to an mbox-ish format
using Outport (outport.sourceforge.net).  I then read the subject and
body of each email and ran a BayesianClassifier.teachMatch("spam",
"...");

I pass it a string consisting of all the words in the subject and
body of the message (separated by a space).  This ended up creating
about 60,000 rows in the word_probablity database.

I wrote a BayesianMatcher class for James (james.apache.org).
Basically,
James (smtp/pop3 server) uses a FetchPop class to pull down emails
from my pop3 account and route those to a local email user.  During this
time I pass the email through to a matcher that uses
BayesianClassifier.classify()
to test whether it gets a 90% or better classification.. if so the
letter
is filed into "deadletters" and I never see it in Outlook.. if its less
than
90% it leaves the email untouched and delivers it.

Problem is.. everything is getting sent to deadletters because of a 0.99
match on everything it received.

Is this not the proper way to use the classifier?  Is there a different
way I should use it to get the results Im looking for?

Any help/suggestions would be appreciated!

Thanks,

- Brent

Re: [Classifier4j-devel] Bayesian Results 0.99 for Everything?

From: Nick L. <ni...@ma...> - 2004-01-03 11:30:01

Are you teaching it non-matches as well as matches? 

It needs something to compare the probability against.

Nick

----- Original Message ----- 
From: "Brent L Johnson" <br...@bj...>
To: <cla...@li...>
Sent: Saturday, January 03, 2004 8:44 AM
Subject: [Classifier4j-devel] Bayesian Results 0.99 for Everything?


> How does the BayesianClassifier differ from a program similar
> to "SpamBayes" (has an Outlook plugin that uses the Bayesian
> algorith to classify emails).  I "taught" the Bayesian classifier by
> teaching
> it with some spam emails I have.  Now it returns a 0.99 classify
> result for practically EVERYTHING.
> 
> A little background..
> 
> I exported about 6400 spam emails from Outlook to an mbox-ish format
> using Outport (outport.sourceforge.net).  I then read the subject and
> body of each email and ran a BayesianClassifier.teachMatch("spam",
> "...");
> 
> I pass it a string consisting of all the words in the subject and
> body of the message (separated by a space).  This ended up creating
> about 60,000 rows in the word_probablity database.
> 
> I wrote a BayesianMatcher class for James (james.apache.org).
> Basically,
> James (smtp/pop3 server) uses a FetchPop class to pull down emails
> from my pop3 account and route those to a local email user.  During this
> time I pass the email through to a matcher that uses
> BayesianClassifier.classify()
> to test whether it gets a 90% or better classification.. if so the
> letter
> is filed into "deadletters" and I never see it in Outlook.. if its less
> than
> 90% it leaves the email untouched and delivers it.
> 
> Problem is.. everything is getting sent to deadletters because of a 0.99
> match on everything it received.
> 
> Is this not the proper way to use the classifier?  Is there a different
> way I should use it to get the results Im looking for?
> 
> Any help/suggestions would be appreciated!
> 
> Thanks,
> 
> - Brent
> 
> 
> 
> -------------------------------------------------------
> This SF.net email is sponsored by: IBM Linux Tutorials.
> Become an expert in LINUX or just sharpen your skills.  Sign up for IBM's
> Free Linux Tutorials.  Learn everything from the bash shell to sys admin.
> Click now! http://ads.osdn.com/?ad_id=1278&alloc_id=3371&op=click
> _______________________________________________
> Classifier4j-devel mailing list
> Cla...@li...
> https://lists.sourceforge.net/lists/listinfo/classifier4j-devel
>

RE: [Classifier4j-devel] Bayesian Results 0.99 for Everything?

From: Brent L J. <br...@bj...> - 2004-01-03 17:34:42

Aaahhhh - that must be the problem then.  I can export some emails
from my inbox and let it classify those as matches and see
if that helps.  And what about categories?  It OK to teachMatch and
teachNonMatch all the emails with the same category name?

Sorry for the simple questions.. but I want to make sure Im using
it right so I get the best possible probabilities for spam
matching at the server :)

Thanks,

- Brent

> -----Original Message-----
> From: cla...@li... 
> [mailto:cla...@li...] On 
> Behalf Of Nick Lothian
> Sent: Saturday, January 03, 2004 6:30 AM
> To: cla...@li...
> Subject: Re: [Classifier4j-devel] Bayesian Results 0.99 for 
> Everything?
> 
> 
> Are you teaching it non-matches as well as matches? 
> 
> It needs something to compare the probability against.
> 
> Nick
> 
> ----- Original Message ----- 
> From: "Brent L Johnson" <br...@bj...>
> To: <cla...@li...>
> Sent: Saturday, January 03, 2004 8:44 AM
> Subject: [Classifier4j-devel] Bayesian Results 0.99 for Everything?
> 
> 
> > How does the BayesianClassifier differ from a program similar to 
> > "SpamBayes" (has an Outlook plugin that uses the Bayesian 
> algorith to 
> > classify emails).  I "taught" the Bayesian classifier by teaching
> > it with some spam emails I have.  Now it returns a 0.99 classify
> > result for practically EVERYTHING.
> > 
> > A little background..
> > 
> > I exported about 6400 spam emails from Outlook to an 
> mbox-ish format 
> > using Outport (outport.sourceforge.net).  I then read the 
> subject and 
> > body of each email and ran a BayesianClassifier.teachMatch("spam",
> > "...");
> > 
> > I pass it a string consisting of all the words in the 
> subject and body 
> > of the message (separated by a space).  This ended up 
> creating about 
> > 60,000 rows in the word_probablity database.
> > 
> > I wrote a BayesianMatcher class for James (james.apache.org). 
> > Basically, James (smtp/pop3 server) uses a FetchPop class 
> to pull down 
> > emails from my pop3 account and route those to a local email user.  
> > During this time I pass the email through to a matcher that uses
> > BayesianClassifier.classify()
> > to test whether it gets a 90% or better classification.. if so the
> > letter
> > is filed into "deadletters" and I never see it in Outlook.. 
> if its less
> > than
> > 90% it leaves the email untouched and delivers it.
> > 
> > Problem is.. everything is getting sent to deadletters because of a 
> > 0.99 match on everything it received.
> > 
> > Is this not the proper way to use the classifier?  Is there a 
> > different way I should use it to get the results Im looking for?
> > 
> > Any help/suggestions would be appreciated!
> > 
> > Thanks,
> > 
> > - Brent
> > 
> > 
> > 
> > -------------------------------------------------------
> > This SF.net email is sponsored by: IBM Linux Tutorials. Become an 
> > expert in LINUX or just sharpen your skills.  Sign up for 
> IBM's Free 
> > Linux Tutorials.  Learn everything from the bash shell to 
> sys admin. 
> > Click now! http://ads.osdn.com/?ad_id=1278&alloc_id=3371&op=click
> > _______________________________________________
> > Classifier4j-devel mailing list 
> > Cla...@li...
> > https://lists.sourceforge.net/lists/listinfo/classifier4j-devel
> > 
> 
> 
> -------------------------------------------------------
> This SF.net email is sponsored by: IBM Linux Tutorials.
> Become an expert in LINUX or just sharpen your skills.  Sign 
> up for IBM's Free Linux Tutorials.  Learn everything from the 
> bash shell to sys admin. Click now! 
> http://ads.osdn.com/?ad_id=1278&alloc_id=3371> &op=click
> 
> _______________________________________________
> 
> Classifier4j-devel mailing list 
> Cla...@li...
> https://lists.sourceforge.net/lists/listinfo/classifier4j-devel
> 
> 
>

Re: [Classifier4j-devel] Bayesian Results 0.99 for Everything?

From: Nick L. <ni...@ma...> - 2004-01-04 08:45:04

Yes, passing the same name will be fine - or else you can pass no category
name and it will just use a "default" category.

Nick

----- Original Message ----- 
From: "Brent L Johnson" <br...@bj...>
To: <cla...@li...>
Sent: Sunday, January 04, 2004 4:04 AM
Subject: RE: [Classifier4j-devel] Bayesian Results 0.99 for Everything?


> Aaahhhh - that must be the problem then.  I can export some emails
> from my inbox and let it classify those as matches and see
> if that helps.  And what about categories?  It OK to teachMatch and
> teachNonMatch all the emails with the same category name?
>
> Sorry for the simple questions.. but I want to make sure Im using
> it right so I get the best possible probabilities for spam
> matching at the server :)
>
> Thanks,
>
> - Brent
>
> > -----Original Message-----
> > From: cla...@li...
> > [mailto:cla...@li...] On
> > Behalf Of Nick Lothian
> > Sent: Saturday, January 03, 2004 6:30 AM
> > To: cla...@li...
> > Subject: Re: [Classifier4j-devel] Bayesian Results 0.99 for
> > Everything?
> >
> >
> > Are you teaching it non-matches as well as matches?
> >
> > It needs something to compare the probability against.
> >
> > Nick