[p2p-hackers] Bounty for Open Source Trust System

coderman coderman at gmail.com
Wed Apr 12 18:00:24 UTC 2006

On 4/11/06, Kevin Walsh <kwalsh at cs.cornell.edu> wrote:
> ...
> One of the earliest wish-list items was to allow more specific voting.
> Maybe I am out of touch, but I was pretty surprised at how many people
> wanted to be able to say things like "The file name is bogus, but the
> bitrate, artist, and file type are all correct." Our most recent release
> has a pretty general framework already in place to handle aribtrary
> statements of this sort.

that seems reasonable; my wife frequently comes across music that has
the wrong artist or title but sounds good and is worth keeping.  in
such a situation it would be nice to vote/rate a subset of the meta
data individually, so the correct parts can be propagated while the
incorrect bits are deprecated and replaced with accurate details.

> The user interface can now generate statements about file types, bitrates,
> and file names, and I don't see any reason not to add other things too.
> The main issue is trying to keep the GUI simple, and being careful about
> the schema. Details are in our nsdi paper due out in a few weeks.

excellent; can you post an update here when it is available?

the user interface issues are usually the crux of the problem,
although a good interface can make explicit feedback useful and
commonly used. (you understand my fondness of implicit metrics, good
UI is not my forte :)

> I guess the problem would be to define "relevant". In the existing
> networks, queries tend to be short, vague, and have no context. Since it
> is not at  all obvious what the user is looking for in the first place, it
> would be kind of hard to decide what is "relevant" in the file sharing
> world.

true.  i should clarify that feedbackfs tracks user ID and program
path (/bin/ls, /usr/bin/firefox, etc) so the relevance of a resource
can vary greatly depending on the application and user.  i tend to
think of recommendation and relevance in a richer context where you
have sequences of resources with detailed implicit metrics attached in
distinct domains of usage (music, web documents, video, etc).  this is
a long term goal and different than simple keyword based searching
where accurate meta data alone can provide relevant results in most

> I'm not too sure what Philippe's bounty is looking for, though. He doesn't
> mention files, or sharing, but does mention "family functions on a global
> level". Can someone clue me in to what that is?

i'd like some more explanation as well.  trust and reputation covers a
lot of techniques and concepts. :)

More information about the P2p-hackers mailing list