A term used at FaceBook for content moderation is prevalence:
Prevalence (% of views) and pieces of content posted are different units. Some content gets millions of views and some none at all. Prevalence is the best way to understand the problem.
-- Guy Rosen, FB VP of Integrity (trust & safety equiv.)
- Hence, 50 billion daily views over 5 billion daily items, with a Zipf distribution.
- Half of all posts get <2 views (excluding submitter)
- Ultimate reach is predictable relatively early based on initial engagement.
I've got a hunch the actual numbers aren't too far from these numbers (say, factor of ten), for daily exposure. Though my sense is that actual tends significantly lower. For Twitter ~ 10--100x lower than FB.
Note that predicted prevalence + moderation scale gives capacity to vet content. The ability to moderate about 1 million items/day (roughly 2,000 human moderators at 500 items/person-day) means catching virtually all items reaching > 1,000 people, if not substantially fewer, which is a pretty good harm-minimisation capability.
Ironic that a post about disinformation is full of it. It would benefit anyone to look up what a conspiracy is and how many things are technically conspiracies. It's basically a buzzword for anything you don't like that hasn't been proven beyond a reasonable doubt. News flash, anything that is nether proven nor unproven is debatable.
2 comments
[ 4.9 ms ] story [ 14.5 ms ] threadPrevalence (% of views) and pieces of content posted are different units. Some content gets millions of views and some none at all. Prevalence is the best way to understand the problem.
-- Guy Rosen, FB VP of Integrity (trust & safety equiv.)
https://nitter.net/guyro/status/1337493574246535168?s=20
I'd really like to have an idea of typical views-per-post (or tweet) distributions on FB or Twitter? These data don't seem to be available.
E.g., in a single (hour|day|week|month|year), there might be (note these are entirely fictitious numbers):
- 1 post gets >1 billion views
- 10 posts get >100m views
- 100 posts get >10m views
- 1k posts get >1m views
- 10k posts get >100k views
- 100k posts get > 10k views
- 1m posts get > 1k views
- 10m posts get > 100 views
- 100m posts get > 10 views
- 1b posts get > 1 view
A few assumptions:
- FB sees ~5b pieces of content (+/- 10x) daily.
- FB has ~2b DAU
- Views/post follows a Zipf distribution
- Daily time-on-site ~ 1hr, 25 items (~2m20s/post)
- Hence, 50 billion daily views over 5 billion daily items, with a Zipf distribution.
- Half of all posts get <2 views (excluding submitter)
- Ultimate reach is predictable relatively early based on initial engagement.
I've got a hunch the actual numbers aren't too far from these numbers (say, factor of ten), for daily exposure. Though my sense is that actual tends significantly lower. For Twitter ~ 10--100x lower than FB.
Note that predicted prevalence + moderation scale gives capacity to vet content. The ability to moderate about 1 million items/day (roughly 2,000 human moderators at 500 items/person-day) means catching virtually all items reaching > 1,000 people, if not substantially fewer, which is a pretty good harm-minimisation capability.