trust flow-downs

[jay]

I waded through the congestion and looked at ciphergoth's trust metric applied to LJ friends-lists.

I'll spare the specific details... overall, about a quarter were communities (I don't read them, as a rule), a quarter were people previously unknown to me, a quarter are familiar but not on my list, and a quarter were people that had once been on my list, but they or I had de-friended the other in the past. Two of those are blocked outright from commenting in my journal, so the algorithm seems capable of picking up enemies as well as friends (or potential friends)... a flaw in a "trust metric", IMO.

Comments

[geekchick]

Well, I haven't looked at how exactly it calculates these things (other than knowing that it looks at friends lists), but I suspect it's somewhat similar to the "popular with your friends" feature; if lots of people you list as friends list these people you're referring to as friends themselves, then it's not surprising they'd show up in your list. It's not like the script has any way of knowing that you've banned people, or that you had them on your list and dropped them.

[brian1789.livejournal.com]

Right... hence my comment back to the author on his page. The closer-in nodes are likely to be more variable, trust-wise, because of greater exposure and experience. He could improve is model by putting variable trust-receptivities on each arc, rather than assuming equal flow-down.