MeetingFribourg

21 - 24 Sept 2004

Fribourg (Switzerland)

Present: paolo, hassan, ...

= local vs global trust metrics =

Local trust metrics computes other peers' trust score based on who you trust/distrust. They take into account subjectivity in trust: I may like peer A, you may dislike peer A.

Global trust metrics computes a global average or reputation (on average, cnn.com's reputation is 9/10). Pagerank is a global trust metric.

Are there any general conditions on when does it make sense to use local metrics, and when global? Perhaps compare time to do extra local information - if "small enough" compared to advantage you get, local makes sense.

Global danger: tyranny of majority, stomping local tastes, or those who are "ahead of the game". Maybe ideal is to always have weighted interpolation of both? To test: i) leave-one-out, then compare local vs global metrics as predictors for left-out-link, local better; ii) (thought experiment) compare relative performances of different settings along slider for a task like "summarizing a field of research" or "open-source (political) intelligence".

Paolo claims local almost always better, if practical. Difficulty is ethical / social subgroupings, Daily Me - how keep any cross-group understanding? Maybe have some cyber-right of reply, i.e. anyone who speaks against me must include a link to my point of view.

http://www.caterina.net/archive/000696.html "a political scientist at the University of Michigan has done some experiment using computer-simulated problem-solving agents that demonstrated that a group that consists of all smart agents does worse than a group that consists of some smart and some not so smart agents."

For sure there are qualities that nearly everyone agrees upon - like 'not killing others', 'not stealing' etc. With those qualities the global measures are the most effective because they decreas the cost of evaluating trust. You just trust the law system and the government, you don't need to evaluate all the trust factors for all your trivial daily transactions (like for example buying something on a street corner). This is the argument I use for global trust measures on my wiki. Of course when we talk about qualities where there is no such universal agreement than this is another strory. -- ZbigniewLukasiak

hi zbigniew. The fact is that even with topics such as "not killing others", that are agreed by the majority of people in the world, there could always be someone that does not agree with such rules and what? Should she be forced to submit the opinions of majority? I don't think so. I think technologies allow to see the world from your point of view, no matter how much different. Anyway, even with "not killing others", are you sure we all agree? [Read: "is abortion ok?"]. The other point is that, in order to society to move on, minority positions must have a way to become majority, otherwise we would still be here burning women randomly because the "majority" thinks they are witches. Did I convey my point? (that is "there is no reality and no global consensus on anything. can technologies cope with this new vision that is possible today and not years ago?").

demonstrate that local make sense with numerical experiments
Can we demonstate at least that there are "controversial users" (trusted by many, distrusted by many) for which a global trust metric does not make sense?

Canonical example might be bimodal distribution, e.g. opposite political inclinations.

simple detect-controversial-users alg
Rank {U} by e.g. the ratio Like(U_i) / disLike(U_i). Ratios "close" to 1 are controversial (since have significant subgroups who like and dislike them).

Difficulties: Define thresholds. Account for indirect trust.

localize pagerank
maybe think of 2 stage process - first stage where computer web of trust or global pagerank (reps for everyone). 2nd stage - add on simple personalized algorithm to filter / scale, e.g. "Personal Google" http://labs.google.com/personalized/

there are some papers

An Analytical Comparison of Approaches to Personalizing PageRank www.stanford.edu/~taherh/papers/comparison.pdf - this paper compares 3 different techniques for personalizing pagerank

Scaling Personalized Web Search www2003.org/cdrom/papers/ refereed/p185/html/p185-jeh.html

In Stanford there is the "Stanford Personalized PageRank Project" http://nlp.stanford.edu/projects/pagerank.shtml

is local more attack resistant?
how can trust metrics be attacked?

attack by clique --  drop old IDs when cheating (defense - link to real ID, low initial trust)  --  profile cloning, then skewing / add spam object (defense w web of trust). Could copy web of trust too, but should be inbound links that matter there!

how can they be attack-resistant?

I reviewed many papers in these last days and all of them make the assumption that there are some "bad" peers/users and the goal of the system is to spot out this globally bad peers. This assumption does not map with reality!

Objective vs subjective...constant tension between shared judgements and idiosyncrasies which may not map to reality. Can only tell globally bad peer when agreed standard of bad behavior in system.

Review Eigentrust and similar P2P reps.

scalability issues?
Personalizing is expensive but useful. Easy start is to have small number of distinct subgroups, each with their own view, then get some weighted average of these based on your own preference for each subgroup. (Blogs are like this in a way - find strongly-connected cliques.) Probably in many common areas people are similar enough that this works.

= trust-aware recommender systems =

my phd thesis

in general, where trust data can be used? for doing what?

Obviously application areas should be i) not too private, ii) not spoofable. Recommendations through web-of-trust - movies, music, places to go, books. Extension of personal social capital network - implicit introductions in social network, or small favors for those close in web-of-trust. Web-of-distrust?

Trust competition testbed
http://www.lips.utexas.edu/~kfullam/competition/

Very similar to robocup, they want to create a model in which different players can play the game. Evaluation metrics will be fixed: "who win the game?" "for doing what?" (in football, it is easier, the winner is the one that scores more goals). Rules of the game will alo be fixed (just as in football, players can kick the ball and look for the ball, in this competition players will be able to "do something" like interact with others, communicate to them their beliefs, sense the world, ...)

it is an open research project waiting for inputs and ideas.

One possibility is an attack - defence on social network game. (See "How to Subvert a Democracy" below.) Some players want to find the truth, others want to hide it...or some want to promote one belief, others another belief. Game board is the social network itself. Moves would be make a link to a new node, push some idea down some links, change trust values, or independently check truth of assertion (this should be a lot more expensive in time!).

= Protocol and Model for "Network of Friends" =

FOAF and semantic web
i was at the foaf workshop 2 weeks ago (see http://sw.deri.ie/~jbreslin/foaf-galway/). it was very interesting

it seems to be the most promising attempt to put in electronic format the "reputation society". there were also some discussions about trust and foaf (see http://rdfweb.org/topic/FoafAndTrust_2fWorkshopNotes)

other proposals are RVW (RDF format for reviews and opinions on things) or OpenReview

privacy and security aspects to storing this info in electronic format?

Trust vs Similarity
They are not the same! Social similarity or shared interest is likely correlated with trust. But combining the two make it hard to separate effects. It also has subtle effect of making those who are less like you seem less "trustworthy", and we already have enough of that...

So take them as different concepts. Trust should be a measure of honesty or reliability. As such it might also include some measure of competence and knowledge. Similarity is a measure of shared interests.

(In fact, trust can easily be made a scalar value, similarity perhaps less so. Also maybe a network of trust needs to be more distributed than a network of similarity, e.g. the latter could be inferred automatically from observations of your implicit ratings through keywords or which sites you go to, but the former is a personal decision and one that's important to keep under personal control.)

I'd like to link to people with high trust but low similarity, just to remind myself what I'm missing.

Polarization Model
One general issue with making finding similar people easier is the "Daily Me" / echo chamber effect.

Potential model: network of people with some distribution of interests. People have some distribution of preferences to trust those who are similar to them. Then look to see under which conditions relatively isolated subgroups form, vs conditions for well-mixed open society.

Simple first idea:

* Distribution of interests. Vector of properties, random amounts of each.

* Similarity: Sum or dot product of two people's interest vectors.

* Trust: Each person has trustThreshold:  if similarity (P_1, P_2) > trustThreshold(P_1), then P_1 is more likely to trust P_2.

Then look at distributions for trustThreshold to see which distributions lead to separate groups.

Can also take into account social structure. Maybe if you are a neighbor (in the social graph) with someone, then you are also more likely to trust them? But if you're around people you distrust, then you're likely to rewire some of your links, looking for a more favorable environment.

(Schelling's Micromotives and Macrobehavior has a (more basic but still interesting) model of separation of people based on similarity.)

This model can also be incorporated into the next model on Subverting a Network of Trust, by imagining that the supernodes are trying to i) change the interests of others, ii) change trustThreshold. Both are interesting - the first case would be like advertising, the second like a politician who wants to raise the fear level to make others seem more dangerous (sound familiar?...) Then the question becomes, how much influence can supernodes wield in these cases?

Experiment: Subverting a Network of Trust
The point of this model is to test how much subversion of nodes is necessary to change a whole population's opinion. There are two cases: Supernodes, and No Supernodes.

(See also "How to Subvert a Democracy" section, later in this document.)

Case: Supernodes
Start off with a NoT. (Try both real and synthetic.) Suppose for simplicity that there are just two opinions, T and F ("True" and "False"), and everybody in NoT starts off believing T.

Add [1 | few] "supernodes", S. Every node in S believes F.

Now every user in NoT adds S as trusted nodes. So now everybody trusts the supernodes (think mass media or trusted authority figures...).

Experiments:

* Vary the amount of trust people have in S from 0 to 1 (x-axis). Graph against the % of nodes in NoT who believe F - i.e. who have been "subverted" to the opinion of S.

* Gradually increase size of S. For each increased size, rerun first experiment. How fast does change happen? (Can combine this and previous experiment in a 3D graph.)

* (Vary the amount or speed opinions are propagated.)

* Instead of all supernodes believing and telling F, what happens if there is some fraction who also tells T?

* Having some supernodes which believe T is particularly interesting if each node in NoT can choose which supernodes it listens to. Then you get into modeling polarizing too (and maybe the polarization model of previous section becomes relevant).

Case: No Supernodes
An alternative to having supernodes who subvert the other nodes would be to try a constant probability P_F that any node will be subverted on a given turn. (One might also want to then give each node a max lifetime, otherwise every node will eventually get subverted no matter what P_F is.)

Then run a simulation where in each round, some nodes are subverted, and all nodes propagate their opinions. Similar to supernode case, would be interesting to graph value of P_F and node lifetime against % of nodes who believe F.

quick comments by paolo (move them around)
It seems very much like a model of virus spreading (ex: aids). At the beginning everyone has not the virus. Then someone with AIDS enter the system and it is someone that has a lot of sexual contacts (s/he is very beautiful and admired) [these are the media: loved and terribly ill inside]. how fast sane people get infected by infectiuos people (the media)?

I have finished "function and structure of complex networks". fascinating and there are a lot of discussion about diffusion of information and viruses on networks. [we may want to reread it in deep to see if something similar was already done]

An interesting related concept is Information cascades (i think studied in psychology and sociology, google for ti)

how can you vaccinate part of the population? the best way seems to vaccinate the media but, assuming this is not possible, you must try to divert edges to alternative media (normal people that are not infectious). Are they bloggers?

--paolo

--

= other points =

privacy, identity
they are cornerstone and basic issues in the "reputation society" but i would prefer to touch them just marginally because i have no clever proposal up to now.

european project
could we consider the idea of submitting a project proposal to EU about "reputation society"?

...a start would be compiling a list of those active in trust / reputation.

suggested papers
(just the ones that come to my mind now ...)

* http://www.cs.ucl.ac.uk/staff/F.AbdulRahman/docs/#phdwork phd thesis in progess of Alfarez Abdul Rahman

For now, there are 2 draft survey chapters that may be of interest:

* Survey of Trust Research in the Social Sciences, from economics, philosophy, political science, sociology and psychology. * Survey of Trust Models for Computer Networks, which compares the current models against findings in the Soc Science survey chapter above

* http://portal.acm.org/citation.cfm?id=503376.503456 Finding others online: reputation systems for social online spaces (mycrosoft research group)

* http://jasss.soc.surrey.ac.uk/7/3/reviews/squazzoni.html "Reputation in Artificial Societies: Social Beliefs for Social Order" - I guess this is a relevant book for you!

jung
http://moloko.itc.it/paoloblog/archives/2004/04/06/jung_java_universal_networkgraph_framework.html JUNG — the Java Universal Network/Graph Framework--is a software library that provides a common and extendible language for the modeling, analysis, and visualization of data that can be represented as a graph or network. It is written in Java, which allows JUNG-based applications to make use of the extensive built-in capabilities of the Java API, as well as those of other existing third-party Java libraries. The current distribution of JUNG includes implementations of a number of algorithms from graph theory, data mining, and social network analysis, such as routines for clustering, decomposition, optimization, random graph generation, statistical analysis, and calculation of network distances (Dijkstra Shortest Path), flows, and importance measures (centrality, PageRank, HITS, Random Walk, etc.). JUNG also provides a visualization framework that makes it easy to construct tools for the interactive exploration of network data. Users can use one of the layout algorithms provided, or use the framework to create their own custom layouts. In addition, filtering mechanisms are provided which allow users to focus their attention, or their algorithms, on specific portions of the graph.

How to Subvert a Democracy
numerical analysis on real data about "most efficient way to bribe different institutions to keep a corrupt government in power, while still pretending to be a democracy." in particular, which is the category most expensive to bribe/corrupt? politicians? lawyers, constitutionalists? ... just answer before going to look for the paper! ;-)

http://blogs.law.harvard.edu/ethan/2004/08/16#a298

This raises interesting general questions about quantifying "robustness" of a social network or web of trust. From a marketer's point of view, they might want to know which nodes are most influential, to best spread their ideas - there have been papers on this, like Kleinberg's:

http://www.cs.cornell.edu/home/kleinber/kdd03-inf.pdf

From a democracy point of view, what effect do the few "mass media nodes" have that a majority trust? When is a web of trust more or less vulnerable to subversion? Could one check automatically the flow through web of trust and look for cut-points, e.g. minimal set of nodes that would i) disconnect graph, ii) be sources of incoming trust for maximal set of other nodes?

conferences
Beyond Personalization 2005 - A Workshop on the Next Stage of Recommender Systems Research (San Diego, Jan 9, 2005) 	 http://www.grouplens.org/beyond2005/

iTrust 2005 (Paris, May 23-26, 2005) http://www-rocq.inria.fr/arles/events/iTrust2005/

AAMAS 2005 (Utrecht, July 25-29, 2005) http://www.aamas2005.nl/

this and others at http://moloko.itc.it/trustmetricswiki/moin.cgi/TrustRelatedConferences

Citizen Deliberative Councils
http://www.corante.com/many/archives/2004/09/18/citizen_deliberative_councils.php

Papers about recommender systems
PapersAboutRecommenderSystems