Trust metric

From Trust metrics wiki - Trustlet, a free, collaborative project for collecting and analyzing information about trust metrics.
(Redirected from Trust metrics)
Jump to: navigation, search

A trust metric is a technique for predicting how much a certain user can be trusted by the other users.

The basic assumption of trust metrics is that trust can be propagated in some way. The reason is that you trust your friend more than a stranger and so, under certain conditions, a friend of your friend is possibly more trustworthy than a random stranger (at least because your friend trusts her).

For example, in the context of movies tastes, if all your 10 trusted cinema-friends trust "mary" (that you don't know) about her movies tastes, can we predict that you will find her opinions about movies useful as well?

A simple example. Suppose you have the following information: ME trusts Alice, Me trusts Bob, Alice trusts Carol, Bob trusts Carol (this is a simple trust network). You might want to ask yourself: should ME trust Carol? A trust metric is nothing but a tool for finding an answer to this question.

Trust metrics can either be local, if they compute personalized values of trust, or global. A global trust metric is a trust metric where trust values for nodes are not personalized, i.e. they do not depend on your own position in the network. With local trust metric, the trust values you see for other nodes depend on your own position in the network. A local trust metric predicts trust scores that are personalized from the point of view of every single user. For example a local trust metric might predict "Alice should trust Carol as 0.9" and "Bob should trust Carol as 0.1", or more formally trust(A,C)=0.9 and trust(B,C)=0.1. Currently most trust metrics are global, but we think that local trust metrics will lead to more interesting applications.

TODO: make this page more precise and less fuzzy

Other possible questions a trust metric can find an answer for are:

  • Trust depends on the context (I trust Bob the mechanic to fix my car but not to look over my son,this is called Specific_trust or context dependent trust).
  • You can add weights to every trust statement (I trust Mena as 9 out of 10)
  • Is trust transitive? How much? In every context?
  • Trust in general is subjective (I trust IndyMedia, you trust CNN) so global trust values (for example the one computed by Google with PageRank) usually don't apply but are usually less computational expensive. When are global predicted trust values not enough?
  • How can I come to know that Mena trusts Cory? Gossiping? Broadcast? DHT? Asking Mena directly?
  • How can I be sure that the information "Mena trusts Cory" was really created by Mena? Here public key cryptography comes into play ... a very interesting topic, but for now we can just assume that all information we receive is reliable because it is digitally signed.
  • The setting (people/trust connections) is very similar to (web sites/links): what are the similarities? what are the differences?
  • Can trust metrics help in detecting malicious users? (rhetorical, this is in fact one of the main advantages of trust metrics)
  • As long as trust values are just total or missing (I trust Alice and Bob, I express no trust statement to all the other people), it is pretty easy to imagine a trust metric. What happens when you can have real value trust statements (I trust Bob as 8/10 and Spammer as 0/10 and Mary as 9.95/10)?
  • What does a negative trust value really represent?
  • Are users willing to give negative ratings to other users (some papers say that the negative ratings in eBay are really a few [xxx find the paper])
  • Do negative trust values just stop the trust chain or I can use them to infer something such as "I should not trust this guy"?
  • How can users easily provide trust statements? What are the best ways to explain it? (Some reports say that eBay users find too complicated the very very simple metric used now, it can be an error to introduce a more sophisticated but more complicated metric!)
  • What representation of trust statements best facilitate users? What about explanation of the results ("You should trust "Mena" as 9.4/10 because ..."
  • What would be a better representation and Visualization of the social network? (I like a lot, but many people say that these kind of interfaces are just very intriguing and then people don't use them because they are too complicated and not very handy)
  • Trust metrics can help in detecting paradox. "I trust Controversial, I trust Hanna, Hanna totally distrusts Controversial". Should I be warned of this? What if *all* my "friends" distrust him?
  • Sociological concerns (Do you trust all your friends? What if a friend of yours sees that she is not in your "web of trust"?)
  • Add a new one! (just free your mind and write down the most unconfessed questions you always had abut trust metrics!! ;-)

See Trust metric input output

We collect all the trust metrics we are aware of at Analyzed trust metrics.

See Where trust metrics are useful.



Trust metric FAQ

Basic questions

What is a trust metric?

How is it used?

Trust metrics are used to model social situations where individual people in the network can't be trusted.

Why is it used?

Where is it used?

Also see Wikipedia

Personal tools