[expertfinder-dev] Re: Expertise and Trust Vocabs

Daniel Olmedilla olmedilla at l3s.de
Mon Jul 9 14:11:07 BST 2007 

Hi to all,

my name is Daniel Olmedilla and since I joined the list recently and this is 
the first e-mail I read on the list, I apologize in advance if some of the 
issues I add have already been discussed.

> > but what about: "i rather trust him..." or "i would probably
> > double-check, however I always made good experiences so far..." or "my
> > friends trust him, so I would also go with him..."
> >
> > in my opinion these are no digital trust expressions, i.e. i could not
> > say: YES or NO concerning the trust i have.
>
> But could you really give a percentage?
> The relations you proposa are interesting, valid examples, but also all
> of them are boolean rather than "fuzzy" or probabilitstic.
>
> > other example: i trust the swiss railroad to be on time in over 95%,
> > however i would only trust the austrian railroad company to 80%. i dont
> > fully trust either of them, but i probably preferred the swiss company
> > ;-)
>
> My point is that these numbers are usually arbitrary (ok, maybe for the
> particular example with the trains, they could be backed by some
> statistics). If you can convincingly tell me how you'd guess the number
> for  "I trust Reto by 73.456% that he knows about RDF", I'd be happy to
> know... (Do you keep a personal statistics on how often your friends say
> nonense about a specific topic and keep this up-to-date? :-) )
> Seriously, I am just not sure how to arrive at such values in a useful
> manner. That's why I said, either saying yes or no
> could be good enough for getting a start.

I agree with both. Trust is not necessarily boolean but a quantitative 
property. Axel is also right in the sense that it is not easy to make up 
explicitly those numbers from scratch. There are two main approaches:

- The user does not define those values explicitly but they are inferred. For 
example, if I define with 1 the train being in time and with 0 it being 
delayed, I could apply some computation to find out my current level of trust 
to the train company. See the Eigentrust algorithm as a nice example of this.

- The user define explicitly the values (e.g., via qualitative tags to 
ranges). For example, I trust completely a person could be a 1 (or 100%), 
while I partially trust someone could be assigned a 0.5, I don't know someone 
would be 0 and I distrust someone is -1. Notice that there is a lot of 
research on this, including discussions over the usefulness of "distrust". In 
social networks,  hybrids approaches are followed, therefore algorithms are 
used in order to transitively follow explicit relationships in order to find 
implicit trust relationships. Jennifer Golbeck has quite some work on this 
are.

> > anyway, also of what i remember from my security classes at university,
> > there are trust evaluation algorithms that are based on the combination
> > of trust values from different parties, arent there?
>
> Not aware, maybe Daniel could post some relevant pointers?
>
> anyway, my claim was that the numeric values seem to be arbitrary and
> hard to assess and I was suggesting to start simple.
>   Also other wider used vocabs have some arguable properties in favor of
> keeping things simple than being over-precice. Take for instance:
>
>    foaf:based_near
>
> In fact this also should be rather a ternary relation which also gives
> the time range (since) when a person was/is based near a certain
> location, but it seems a simple and possibly incomplete way to represent
> information was chosen dliberately here.

In trust you also may consider context a relevant issue. I trust Axel on 
logic-related topics but not on medical issues or piloting a plane.

Cheers,

	D.

>
> I'd at least guess we should have means for both, a very simple and a
> more fine-grained vocabulary.
>
> best,
> axel
>
> > best regards,
> >
> > reto
> >
> > Axel Polleres wrote:
> >>Re: my mail form 27/06 ....
> >>
> >>I'd like to try to solicit again comments from your side on my
> >>comment on the numeric-valued ratings of trust in hoonoh and the trust
> >>ontology especially on 1)... I really don't think it is a good idea to
> >>"rate" trust, I'd rather say something like I trust somebody on
> >>something, but not I trust somebody with a degree of X...
> >>
> >>comments?
> >>
> >>axel
> >>
> >>Axel Polleres wrote:
> >>[...]
> >>
> >>>1) I would rather only go for "boolean" ratings at all.
> >>>   You trust someone or not. Or maybe, if you want it more refined,
> >>>   you trust someone on something. But it leaves too much ambiguity to
> >>>   say I trust someone with a degree of X.
> >>>
> >>>   In doubt, I would rather suggest to go for the simplest option.
> >>>
> >>>   Ranks could rather be *derived* based on "how many" trust or based
> >>>   on some other ranking function aggregating networks of trust... but
> >>>   this is not an issue for a vocabulary but rather for algorithms on
> >>>   top (ie. "enabling technologies" as we called it in previous
> >>>   discussions)
> >>>
> >>>2) I'm unsure whether it is a good idea to superclass foaf:properties:
> >>>     "foaf:knows might be a subproperty of trust:trust0"
> >>>   people might have a different understanding, of what knows means.
> >>>   I could know someone and this doesn't say anything about how much or
> >>>   little I trust this person. If even, I would put it the other way
> >>>   around, but also that sounds fishy. I could trust the dalai lama,
> >>>   although I don't know him personally
> >>
> >>[...]
> >>
> >>>>I'm not sure if you've seen my draft Hoonoh ontology Andreas. It takes
> >>>>a slightly different approach, but fulfils the same kind of role. For
> >>>>example, using the ExpertiseRelationship class and associated
> >>>>properties one can describe an expertise relationship between a
> >>>>foaf:Person and a Topic (hoonoh:Topic, subclass of skos:Concept), and
> >>>>assign this relationship a value.
> >>>
> >>>See above, I have the same concern with the numeric value here as above.
> >>>I personally wouldn't "rank" my levels of trust". At max I would make
> >>>some comparative statements, e.g. I" I trust X more than Y" (which by
> >>>the way I wouldn't make public, though ;-) )
> >>>
> >>>> I hope that it's a relatively clean
> >>>>bit of modelling, but would be pleased to get feedback.
> >>>
> >>>My feedback for a small starting vocab would be to keep it simple
> >>>as simple as possible/necessary, as I said.
> >>>
> >>>>What we're doing is generating metrics for these
> >>>>ExpertiseRelationships (and ExperienceRelationships, and
> >>>>ImpartialityRelationships) linking people to Topics, and two others
> >>>>that link people to people (AffinityRelationship and
> >>>>TrackRecordRelationship). By following this approach we can publish
> >>>>one "opinion" (that generated by our algorithms) about someone's
> >>>>expertise/experience/etc in a particular topic, but leave individual
> >>>>applications to decide whether that person is trustworthy in a
> >>>>particular context, and depending on the user needs (i.e. people's
> >>>>thresholds for trustworthiness will defer).
> >>>>
> >>>>>URIs for various skills & tags can be provided by dbpedia
> >>>
> >>>On the expertfinder wiki, we suggested that
> >>>  ACM classifications or maybe also the subject areas at the WWW
> >>>  Conferences Archive could be interesting. Also wikipedia terms (as
> >>>used above, were a suggestion which came up several times already).
> >>>see
> >>> http://www.rdfweb.org/topic/ExpertFinder_2fBestPractices
> >>>
> >>>If somoene wants to add/comment there, feel free!!!
> >>>
> >>>BTW: As a concrete exercise: What would be a good uri for
> >>>Trust&Security? see the above example.
> >>>
> >>>>Agreed. This is what we're planning, or at least topics that map to
> >>>>dbpedia if possible.
> >>>>
> >>>><snip/>
> >>>>
> >>>>>@tom, when do we start the website "Sparql query of the day"?
> >>>>>
> >>>>:D As soon as possible! Would contacting SWEO be a good strategy?
> >>>>
> >>>>Uldis is a member of SWEO AFAIK and may be able to present the
> >>>>proposal to that group. I imagine many of the members there would be
> >>>>interested in the idea, and their contributions/backing would be very
> >>>>helpful.
> >>>
> >>>not yet sure whether i understand the sparql of the day idea really,
> >>>but sounds like fun! :-)
> >>>
> >>>
> >>>axel
> >>>
> >>>>[1]
> >>>>http://lists.foaf-project.org/pipermail/expertfinder-dev/2007-June/0002
> >>>>09.html
> >>>>
> >>>>[2] http://kmi.open.ac.uk/people/tom/tmp/draft-hoonoh-ontology.rdf
> >>>>[3] http://kmi.open.ac.uk/people/tom/tmp/sample-hoonoh-instances.rdf



-- 
     Dr. Daniel Olmedilla
     L3S Research Center and Hannover University
     Appelstr. 9a
     D - 30167 Hannover

     Phone: +49 (0)511-762.17741
     Fax:   +49 (0)511-762.17779

     http://www.L3S.de/~olmedilla/
     E-Mail: olmedilla at L3S.de

More information about the expertfinder-dev mailing list