Meeting 204 20august 202004

Brief summary of 29 July 2004 Meeting

Topic: Trust metrics

Presents: Marco Gori Paolo Avesani Paolo Massa

= Short Introduction =

web and trust networks: differences.

in the web (as a graph): sites are nodes and links are edges.

in a trust network (as a graph): sites are people (or peers in a p2p net, or cooperating servers or mobile devices or ...) and links are trust statements.

Trust statement are weighted (in [0,1]) while links are not.

A--(1)-->B : means A trusts B (as 1)

A--(0)-->C : means A trusts C (as 0), that is "A distrusts C"

A missing edge D : means A don't know directly D and is hence not able to express a trust statement

Paolo--(0.8)-->Marco : means Paolo trusts Marco as 0.8 (quite high but not total! ;-)

Are trust statements used? Yes. In many sites, p2p networks, on the semantic web (FOAF format!).

See more on the presentation I made few days ago in Trento Univ:

http://dit.unitn.it/research/seminario?id=03-03-016 (abstract)

http://ardent.unitn.it/seminars/slides/Massa.pdf (PDF slides)

(introductive talk without very deep concepts or experiments)

--

My questions to Marco:

= are there search algorithms that use the concept of negative links (similar to distrust statements)? =

on the web it is not possible to express "negative" links: in fact PageRank consider every link of site A to site B as a positive vote of A about B. But, consider this: this guy is stupid In this case, the link is a "negative" vote.

Questions: are there search algorithms that use the concept of negative links (similar to distrust statements)? are there papers? has pagerank be reformulated in order to take into account this? Would it be possible to reformulate Pagerank in order to consider this? Is it enough to put 0 (or -1?) in the matrix instead of 1 for the distrust statements? would everything work? is it interesting to think about it? Is it possible to use PageRank with weighted trust statements (in [0,1]) and everything will still work (convergence, meaningfulness of final matrix)?

example of weighted input matrix for the reformulated pagerank paolo	marco	maria	giovanna paolo 		0.8		1 marco	0.9			0.1 ...

Answer of Marco: There are no search algorithm exploiting the concept of negative link, to the best of my knowledge.

are papers trying to understand if a link in the graph structure is a positive or negative link?
This can happen on the web (see previous example) but also on bibliometrics (for example, when in my paper i cite your work as wrong or incomplete or weak)

= are local trust metrics used in search algorithms? =

Global trust metrics and Local Trust metrics: for every user global trust metrics compute a global "reputation" value (for instance, PageRank is a global trust metric). Instead, local trust metrics take into account, for every single user, her subjective views and predict personalized trust values in other users (i could trust George Bush and you could distrust George Bush and this is totally normal).

example: A--(1)-->B--(1)-->Z A--(1)-->C--(0)-->Z

a global trust metric will tell that Z is authoritative as 0.5 a local trust metric will tell to A that Z is trustable (as 1) to B that Z is distrustable (as 0)!

Question: are local trust metrics used in search algorithms? personalized search engines?

Answer of Marco: i don't think so.

Suggested paper by paolomassa: " Trust Management for the Semantic Web" by Matthew Richardson, Rakesh Agrawal, and Pedro Domingos. it propose a local version of pagerank (basically, the random walk jumps back to the starting node (ME!) with a probability p>0) read my comments at http://moloko.itc.it/paoloblog/archives/2003/11/14/trust_management_for_the_semantic_web.html

About personalized search engines, You should check http://www.eurekster.com/ (results based on what your friends are searching and clicking) but I think it is not a research project. A9.com (by amazon) has somthing similar (but i'm not sure).

= there are some papers that analyze the behaviour of pagerank (or other ranking algorithms) on controversial pages? =

Question: there are some papers that analyze the behaviour of pagerank (or other ranking algorithms) on controversial pages. something about "second eigenvector of pagerank"?

Answer of Marco: no. I was wrong about Spectral Analysis and Marco briefly explained it.

-

= Other points =

* Paolo Massa: there are no paper evaluating different trust metrics but it is a very active and recent topic!

* You may want to check the wiki I set up  "trust metrics evaluation project" http://moloko.itc.it/trustmetricswiki/moin.cgi/FrontPage

* for graph analysis, I use JUNG: (Java Universal Network/Graph Framework) [java open sourece package to model and visualize java. it has already coded pagerank, hits, markov walks and much more!] read my comments at  http://moloko.itc.it/paoloblog/archives/2004/04/06/jung_java_universal_networkgraph_framework.html or visit it at http://jung.sourceforge.net/index.html

--

Something interesting that emerged in the discussion:

1) six degrees of separation means social networks have a diameter of 6 (more or less). instead on the web the diameter seems to be 18/19. this could be because in human networks there are people highly connected (the connectors with thousands of outgoing links) while in the web there aren't pages with thousands of links.

2) Marco suggested the following paper "Difference Equations on Weighted Graphs" by Bensoussan, Menaldi Heavily mathematical ;-) I need to dig into it.

3) Marco introduced his work about learning on graphs (where instantiations of graphs are the training test). PaoloAvesani has read the final paper and will pass it to me.

4) Is "trust metric" a keyword everyone agrees on? No, some people use "reputation systems" (mainly economists). "trust network", "web of trust", "social network" are keyword but they are not agreed and people will always define them in paper. The topic is very new. The first important paper naming "trust metrics" was by Levien R. Levien and A. Aiken, ``Attack resistant trust metrics for public key certification." 1998 and them his work on http://advogato.org that was the test bed for his trust metric: http://advogato.org/trust-metric.html

Did I forgot something? Do you have any question?