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.
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.
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.