Re: GroupLens (gnus-gl.el)

[Raja R Harinath <harinath@cs.umn.edu>]

Hi,

Lars Magne Ingebrigtsen <larsi@gnus.org> writes:

> Raja R Harinath <harinath@cs.umn.edu> writes:
>
>> But, GroupLens is not supposed to be a moderation system.  It is
>> supposed to find people who appear to have opinions similar to yours,
>> and weights the score based on that "closeness" criterion.
>>
>> It is thus less likely to be poisoned -- unless you are poisonous too :-)
>
> It's a more subtle tool than, er, "flat" moderation, but it's still
> pretty easy to poison 

But the subtlety is of the same kind that makes bayesian filtering
work.  The larger your corpus of ratings, the more resilient to
poisoning.  The poisoner has to put in more work to get to you, and
he'll only get a small payoff for all that effort (fewer people
reached for amount of effort expended).

> (unless you use some kind of authorization system, or keep it "in
> the family").  

I think the issues here are

  * make sure that submitted ratings are credited to right person
    (simple PGP signatures)
  * make sure that the database maintainer is trustworthy 
    (unlike Amazon.com ;-)

Bogus data will be easier to ignore, since it'll be hard for that data
to be "close to" good data.

> Hacking up a system that would inject a gazillion fake profiles that
> would make it likely that your important MAKE VIAGRA NOW messages
> rate highly isn't particularly difficult.  Probably.

But, if none of them is "close", that's not a problem.  And, each time
you come near a poisoned profile, you can distance yourself from it
with your own ratings.

- Hari
-- 
Raja R Harinath ------------------------------ harinath@cs.umn.edu