DiscussionWithClaudia

Discussion with Claudia

just click "edit text" at the bottom of this page to edit it

---

> > Sorry for the delay! Lectures restarted in Germany last week and

> > therefore the last weeks were rather chaotic.

> >

> > Thanks for the links! The citeULike network was new for me.

> It looks

> > interesting.

> >

> > In my phd studies, I'm working on trust management systems that

> > support users in organizing their trust statements, to

> include trust

> > data from other users (probably from different networks). The trust

> > statements are intended to be more complex than just

> numerical values.

> > Using "concepts" for trust instead of values, the trust

> metrics won't

> > be any more mathematically as clear as they are now.

> this is not good because trust metricsare not mathematically

> clear at all now! ;-)

I hope that I can do something with inference when I define the concepts

in an ontology. There is also some work on measuring the similarity

between concepts. But that's future work...

>

> >The trust management system can

> > then be used by different application such as a trust-based

> >recommender or a trust-based mail filter etc.

> >

> > That's what I'm planning to do in my PhD. Until now, I've

> tried to see

> > what trust networks are available and how they work. I've compared

> > different trust metrics. That's all more or less on the

> > state-of-the-art.

> wow!

> can you send me tha report?

I don't have a report yet, I've only some notes for myself but it is not

yet integrated in one technical report or something else. For the trust

propagation, I've looked above all at Golbeck (simple path algebraic

trust metric), Ziegler (Spreading activation strategies), Guha (Trust -

Distrust - the only metric that includes distrust), Richardson (includes

statements about documents)

> did you check: http://trust.mindswap.org/cgi-bin/relationshipTable.cgi

Yes, I've crawled and parsed the files. The degree distribution for the

outdegree follows a power-law. But the data is not so good and it is not

based on some concrete interaction within an appliction like in

Epinions. So it is unclear where the data comes from, the ratings are

mostly very good and they seem to me rather like everybody is testing

the tool for creating the files but does not think much about the trust

relations.

> http://www.cs.ucl.ac.uk/staff/F.AbdulRahman/docs/thesis-final.pdf

> (thesis of Alfarez Abdul Rahman's

> " There are two rigorous surveys in it, one on social trust

> and the other compares existing computational trust models

> with social trust properties. There is also a chapter on

> threats to reputation systems - its very much a work in

> progress so comment in this would be great. The proposed

> Ntropi trust model is included also, and it is formally

> specified using Z." ?

I'll look at it ...

>

> >

> > The last weeks, I've looked at the properties of trust

> networks. That

> > means that trust networks have more or less a power-law

> distribution

> > for the indegrees, i.e. the number of persons that trust a certain

> > person, and the outdegrees, i.e. the number of trust

> statements that

> > some person is making.

> right!

> http://www2004.org/proceedings/docs/1p403.pdf

> in this one there are 2 graphs that plot it for epinions.com

> trust networks

I found interesting that they only say that there is a power-law

distribution but they do not explain the differences to a other

power-law distribution like in the web. And I think there must be a

difference between epinions and other trust networks due to all their

mechanisms that influence the users' behavior...

- - Discussion about the evolution of networks:

> > However, I think that most real networks differ in some way from the

> > power-law distribution. I'd find it interesting to analyze why this is

> > the case. For example in epinions, the distributions should be

> > different due to the 'eroyalities' system where you can earn money and

> > all the other mechanisms introduced by epinions. But I have not yet

> > analyzed which effects such mechanisms have on the structure and the

> > properties of the network.

> i'm interested in this. if you want to argue more, that could

> be very interesting. i think that we can analyse the time

> evolution of the network ... when you analyze all the network

> at time (end_of_world), the network is already settled down

> and yes you got only powerlaw (though the exponent is -1.7

> while in general on the web it is below -2)

... Yes, I also think that it would be good to include the evolution of

the network. I've already thought about some basic mechanisms for growth

that will result in networks with power-law distributions for indegree

and outdegree. But it is not finished and not implemented at all.

Would you be interested in explaining why epinions does not follow a

mere power-law distribution but that of the epinion's mechanisms

influence it? What ideas do you have about it?

>

> Open Rating Systems

> http://tap.stanford.edu/wot.pdf

>

> > This is not a basic part of my phd. Perhaps you find it interesting

> > too and we could work on it. > yes yes yes

>

> long review (54 pages!!!) of "The structure and function of

> complex networks" http://arxiv.org/abs/cond-mat/0303516

> with table 2 presenting all the networks studied (at that

> time) with their exponent of power law

>

> > I don't know if this would be easily feasible

> > because I've not yet made any analysis on the epinions network.

> I have ;-)

> last 4 papers are about it (i seem an old professor that got

> a dataset and spent all these days on it ;-(((

I find it very difficult to get other data sets with explicit trust

statements and not only implicit ones like using someone's favorites.

Furthermore, I'm looking for trust and distrust and there are only few

ones. The FOAF files with trust (see Golbeck) would be good for me but

there are not so many.

>

> i submitted one to web intelligence 2005 (at 2 in the night!!)

>

> i think i'm posting it on my weblog today or tomorrow...

> if you like monitor http://moloko.itc.it/paoloblog/

Can you send me a copy?

>

> > I've

> > only crawled and looked at the FOAF trust network built by Jen

> > Golbeck.

> >

--- ---

> > Let me know what you think about it and in what you are

> interested to

> > work on for a paper.

> - economic incentives and how they affect epinions.com

> network? (very hard paper, i don't think i can say something

> meaningful here, do

> you?)

I don't know. It's rather what I've writte above with the epinions

mechanisms. It would be interesting if the structure of the epinions

network differs very much from the structure and distributions of other

networks. Then we could hypothesize that it is due to the economic

incentives and the whole system with advisors etc. Do you know some

system with which it could be compared?

> - time evolution of Epinions.com network (maybe better, for

> example how cliques of friends form in time and how attackers/spammers

> perform...)

Guha started to analyze in one of its papers the different structure of

the spammer components but he did not analyzed it very in detail. Do you

think that we could get some good results?

- - - Summary of the above discussion about the evolution of networks and some comments:

- Trust networks follow a power-law distribution --> Is it interesting to analyze the differences that exist in real trust networks? Could the differences be explained? --> Are there any differences from the "pure" power-law distribution in the epinions data set? Are they due to the economic incentives???

--> I've already worked on a growth model (not yet implemented) but it could constitute a basis for the work

- time evolution of Epinions.com network (maybe better, for example how cliques of friends form in time and how attackers/spammers perform...)

--> Would it be possible to integrate the behaviour of spammers explicitly in the growth model? And distinguish them from the cliques? I think that there have been some analysis on the different structure of clusters of spammers and clusters of friends.

We have an agent-based platform that can simulate the behaviour of agents related to which papers they cite (COM project--> Bamberg). It could be extended to simulate the growth of the trust network, too. Until now, the algorithms for the generation of the trust network are very very basic. But it allows for the integration of 'statistical actors' and BDI agents. We could have a growht model for trust networks in which most actors act according to preferential attachement and some actors (the BDI agents) have a more complicated behaviour, either as friends who directly take their real world friendships in the trust network or as spammers.

--> we could compare two networks that are created with a differently parameterized growht model, one growth model for a "nice" trust network with indegree and outdegree clearly following a power-law and a second growth model that generates a network including spammers

--> comparison of both networks with the epinions network --> can we derive from the network structure that some group is rather a "normal" trust component or a group of spammers?