Trust decays exponentially with distance in the social graph, but it does not immediately fall to zero. People who you second-degree trust are more likely to be trustable than a random person, and then via that process of discovery you can choose to trust that person directly.

Some problems with that argument:

* If the variance is high enough, the average hardly matters. You're just getting slightly better noise. Example: a friend is into a sport, so does that mean you want to watch that sport? A friend watches a lot of Fox News, so does that mean you want to watch it?

* The "people I trust" graph isn't captured by "people I follow." You can follow people you don't trust, for reasons like entertainment or keeping track of what groups you dislike are up to.

* "Trust" is an abstract term that isn't well defined out of context. Trust for what purpose?

* Arguing in terms of an abstract social graph obscures whether we are talking about people or organizations and how we construct any actual graph. People follow others for whimsical reasons and often don't maintain their following list. If you ask them why they are following someone, they might not even know.

For the limited purpose of finding interesting people to follow it can be okay, but I don't see it getting automated in a way that would work for web search or finding people with a common interest. For example, Reddit often works better because you're looking for something that can't be found by asking people you know. The people are out there but you're not connected.

I think trust is somewhat transitive, but it's not domain independent.

I have friends whose movie recommendations I trust but whose restaurant recommendations I don't, and vice versa. I have friend that I trust to be witty but not wise and others the opposite.

A system that tried to model trust would probably need to support tagging people with what kinds of things you trust them in.

This. Nobody is 100% trustworthy in every circumstance. When I say I trust someone, what I mean is that I have a good handle on what sorts of things that person can be trusted about, and what sorts of things they can't.

Exactly - you have a reasonable model of a person. So it also includes things like a recommendations giving you the _opposite_ of the purported opinion. Or trusting the details are technically true, but missing the forest for the trees. Or any other contextual interpretation of the data.

I'm sure everyone says that about themselves. It doesn't work out like that though.

I know many people I like very much… but would never trust their judgment.

On second thought, I'm not even sure what "transitive" means here. It seems like it should mean that if you trust your friend's movie recommendations then you trust your friend's friends' movie recommendations? Or maybe something like:

   trustsMovieRecs(A, B) and trustsMovieRecs(B, C) => trustsMovieRecs(A, C).

Their movie recommendations are likely some function that takes their friends' movie recommendations as input (along with watching them), but that's more like an indirect dependency than a transitive closure.