45 sats \ 6 replies \ @k00b 12 Jul 2022 \ parent \ on: Daily discussion thread
How you earn rewards now:
- Post a top ranking story in the last 36 hours
- Post a top ranking comment in the last 36 hours
- Upvote content that eventually becomes top ranking
Is it possible that the trust algo will end up having too much of a feedback loop?
Example:
Days 1-2:
Those who upvoted stories that became top ranking see their trust level increased.
Day 3:
Those who earned higher trust upvote new posts from day 3, but not all of these trusted users vote on the same posts. And thus because the trust rank for those users was high (from properly predicting the top posts days 1 and 2), the posts they upvoted on day 3 don't rank because for any specific day 3 post there's not enough upvotes from "trusted users" to overtake the posts that ranked at the top days 1 and 2.
In other words, trust earned today is a double edged sword tomorrow, unless everyone trusted upvotes the same way tomorrow. And tomorrow's posts can never outrank today's posts as a result. If it weren't for decay, today's top ranking posts would stay there forever, regardless of voting on later posts.
My suspicion on this is because posts from a while back (e.g., 48 hours) are returning to the front page. If they dropped off the front page because of decay, I would assume they wouldn't return if there was new content getting upvotes. (and, it being Wednesday, there's no shortage of content and users / upvotes).
reply
I'm not seeing what you're implying with trust. Top ranking stories aren't what determine trust - they only determine rewards. Unless I misunderstand what you mean.
Highly upvoted posts are in fact stickier than something new that has a handful of upvotes. This was a recent ranking change because new, low quality stories, as determined by upvotes, were pushing out stuff that was older (but still recent) and high quality as determined by upvotes.
Prior:
upvotes/time^1.3 where upvotes are weighted by trustPresent:
upvotes^1.2/time^1.3 where upvotes are weighted by trust1.2 might be too large, or 1.3 too small.
It sounds like you like more turnover of stories and less discussion?
reply
Top ranking stories aren't what determine trust - they only determine rewards.
Right, but two users upvoting a post that later becomes top ranking increases their trust weight, if I am understanding the following statement correctly:
In this release, trust is gained when two users upvote the same piece of content
SN release: enhanced trust algo and more tangible rewards
#42401
But after seeing the algorithm from your comment above, I think I see where I was misunderstanding things. (Well, my misunderstanding is more based on me not previously spending time trying to understand how rankings and trust weighting really work.) But knowing the answer to this next question will help clarify things for me:
Are the trust weights re-computed in real-time as the request for the ranking occurs, or is the trust weight saved with the upvote at the time the upvote occurs?
Let me use an oversimplified example. Assume this is launch day for SN, and the current ranking algo is in place. There are only five users (A, B, C, D, and X). Only two items exist (posts #101 and #102, both posted by user X), and the upvotes occur chronologically as shown, but in quick succession.
| User | Item #s upvoted |
|---|---|
| A | #101 |
| B | #101 |
| C | #102 |
| D | #102 |
So at this point after these four upvotes, how would the ranking look?
Would item #102 be ranked first and item #101 ranked second? Or would it be the other way around?
I can't say because I don't know when the "trust weight" is computed. Here are the two methods I am thinking could be used, but I don't know which is used.
- Trust Weight Computation: Real-time, when the request needing ranking occurs.
At this point, I presume all four users would have the same trust weight. If so, then I'ld assume item #102 should be ranked in the top spot, as that is the newest content and everything else is ~ equal.
or
- Trust Weight Computation: Saved with the upvote, at the time the upvote occurs.
I would assume that A and B would have a better trust weight because at the time of B's upvote, there were just two users upvoting and both upvoted 1 time and for the same item (item #101) -- i.e., "perfect, 100% in agreement". Later when C upvoted, C's trust weight would be lower because it diverged from A and B. And finally when D upvoted, both C and D's trust weight would rise a bit because they both upvoted for item #102, but that is 2 of 4 upvotes, and thus both C and D would not have a trust weight as great as either A or B. And if this is true, then item #101 would likely be ranked in the top spot, and the reason being that the upvotes were from users with a greater trust rating (assuming time decay in this instance is insignificant).
So my concern was that C & D would end up with a lower trust weight even though their behavior was identical to A & B (i.e., two users, voting similarly), but because their upvotes occur later, they get a lesser trust weighting.
After thinking this through further, I now think that you probably are computing the trust weighting in real-time at the time the request for the ranking occurs, and as a result I just wasted about 5 or 10 minutes of your time.
reply
I compute trust once daily, not in realtime. Trust weighted means that if your trust is .7, your upvote counts for .7, and if your trust is .9, it counts for .9.
Happy to help. It might be better to ask questions though so we can get on the same page quicker.
reply
streaming sats would be better, like we will know when it shifts from 1sat/hour to 10 sats/hour to 100 sats/hour or 1 sat/second stream etc.
reply
That'd be interesting, but I'm not sure it's much better, because you definitely don't want a notification every minute. You'd still probably want only a daily report on how much you earned ... in which case there's really not much point to complicating things by sending them to you every minute.
reply