It adds the label if both the following is true:
- you are not following the contact
- it has a 80% similar name and profile picture as someone you already follow
For picture it checks pixels and for name it uses Jaro-Winkler from https://github.com/autozimu/StringMetric.swift I tried levenshtein before but it seemed a bit buggy. Both could probably be improved but I haven't had false positives for a while, there were a few a while back but that was actually related to code errors, not the algorithms.
Thanks. Do you think “in code” wd be better than “in person” for verification of the same data points?
IMHO a crowdsourced solution (allowing people to “verify” each others names and pfp) allows erroneous claims (from a single person) to be ironed out by the claims from others. I imagine this wd be easier to implement and to maintain and produce better results than an algo (a black box for many) locked in code.
Your thoughts?
(ICYMI that screenshot above is a false positive.)
They could complement each other if we use both. Right now it’s possible to send reports but when receiving reports it doesn’t do anything with them yet. We could leverage WoT to decide which reports are relevant or not.
I think so. We’ll be working on this for an onboarding client. Will circle back with you for integration ideas.
