A quick explanation of TrustFlow

[ciphergoth]

TrustFlow does not look at interests, who reads your journal, or any other such thing. It looks only at "friends" lists. It's trying to determine who is closest to who based on who lists who as a friend

Comments

[ciphergoth]

Yes. However the bigger the bucket the less difference it makes. With a 500 token bucket it should be OK.

The correct solution to this is the continuous version of the trust metric, in which "trust liquid" pours continuously into my bucket, then pours equally into those of all my friends, and so forth. However this involves solving 50 systems of linear equations each with up to 800 variables, and I don't know enough about the subject to do that efficiently.

Anyone who knows a good algorithm for finding the eigenvector with eigenvalue 1 of a somewhat sparse 800x800 matrix, given an OK approximation to that eigenvector, please do get in touch.

[ciphergoth]

All values are real and non-negative. Many are zero. No, it's not symmetric

[ciphergoth]

From there finding the trust for all the untrusted people is just a question of adding it up.

Actually the last step is more complex than that. Finding the rate at which trust liquid flows into the buckets of the untrusted people is just a question of adding it up, but from there you have to take into account the liquid already in the buckets to figure out whose bucket(s) will fill first. You then add the right amount of water to all untrusted buckets, add the newly full bucket(s) to the trusted list, and go around again, recalculating the eigenvector for the newly enhanced graph.

I talked about this a little in the first entry presenting the trust metric.