Re: Summary of interests ..

[email protected] wrote on 02/24/2004 01:12:07 PM:

> > Jesse Dougherty wrote:
> >
> > It's a good idea to have users preferences available to filters when
> > they're doing their job, but users preferences is probably easier
> > handled by the MDA (mail delivery agent), not by the MTA. Otherwise,
> > filters will need to redo all rewritting work already done by
> > the MTA,
> > or be more integrated on MTA work flow.
> 
> 
> Good point, should we add an interaction which includes the MDA?

Agreed, rewritting does make it difficult to do personalization using an 
MTA based filter.   However, I believe personalization is necessary for 
most successful filtering deployments.   This means we either have to move 
the filter down the chain to the MDA or figure out a way to make it work 
at the MTA.  Moving all the filtering to the MDA is not likely since there 
are many advantages to filtering as far upstream as possible, but it may 
be possible to split duties between the MTA and MDA.  However, this 
introduces the computation and maintenance costs of requiring a minimum of 
two antispam servers. 

We have been throwing around a solution to this problem in our group that 
is not great, but I am going to mention it to see if it generates some 
better ideas.  Before I go on, I should also mention I am not an SMTP 
expert, but I am learning fast.   Okay, the idea is to add a header to the 
incoming message indicating the original id specified in the RCPT command. 
 This header is passed along to the client.  Then, when a user provides 
feedback on a message, it pases back the stored RCPT value back to the MTA 
antispam filter.   The MTA can then update its model for the indicated 
RCPT id and use this updated model to process new messages destined for 
the same RCPT id.  If a user receives mail using more than one RCPT 
address, then the antispam filter will personalize each RCPT address 
separately.   While not ideal, this is a good approximation to individual 
preferences.

The main problem with the above is that it assumes that each RCPT address 
maps to a single user.   If a RCPT address is rewritten into multiple 
users, each user's feedback will be applied to the same RCPT address and 
therefore result in a combined preference for the RCPT address rather than 
an individual preference.   This may be acceptable, but I suspect not. The 
only alternative in this instance I think is to apply the rewrite rules 
before filtering.  Sigh...

- Rich

--------------------------------------------------
Richard Segal
Agents & Emergent Phenomena
IBM Research

Email: [email protected]
WWW: http://www.research.ibm.com/people/r/rsegal
--------------------------------------------------