Deriving the Reddit Formula • Hacker News Board

Re: Deriving the Reddit Formula

#11

Post by Fando » Wed, Jul 15, 2015, 4:53 AM UTC

raldi wrote: post #5

Note: The author of this piece wrote the original analysis ( http://www.evanmiller.org/how-not-to-sort-by-average-rating.... ) that directly led to my advocating for reddit's "best" comment sort, including to xkcd's Randall Munroe, who managed to finally convince the rest of Team Reddit to make the change ( http://www.redditblog.com/2009/10/reddits-new-comment-sortin... ). So when the author says, "I realize that pro…

Wow, I didn't know that. I was going to say that a long time ago I randomly read Evan's article about 'how not to sort by average rating'. Good work.

Re: Deriving the Reddit Formula

#12

Post by gohrt » Wed, Jul 15, 2015, 5:03 AM UTC

tedunangst wrote: post #7

jsnell wrote: post #4

Do the conclusions change if we assume that the goal is not to optimize for value to the user, but for value to the company? Specifically, it seems that using the vote difference rather than the vote ratio would help controversial stories rank high. Controversial stories in turn are great at driving lots of discussion, which means lots of visits and revisits to the comment section. (Compare to the apparent HN policy…

Possibly short term, but if you build a site around the kinds of people who like flame wars, your advertisers may notice they're getting very little traction from your site. Page views go up, but uniques go down.

Have you met Reddit? They love flamewars.

Re: Deriving the Reddit Formula

#13

Post by tedunangst » Wed, Jul 15, 2015, 5:18 AM UTC

gohrt wrote: post #12

tedunangst wrote: post #7

Earlier quoted context omitted.

Possibly short term, but if you build a site around the kinds of people who like flame wars, your advertisers may notice they're getting very little traction from your site. Page views go up, but uniques go down.

Have you met Reddit? They love flamewars.

reddit also, as far as I know, fares rather poorly on the $/view metric. (It's not quite that simple, sure.)

Re: Deriving the Reddit Formula

#14

Post by Natsu » Wed, Jul 15, 2015, 5:21 AM UTC

gohrt wrote: post #12

tedunangst wrote: post #7

Earlier quoted context omitted.

Possibly short term, but if you build a site around the kinds of people who like flame wars, your advertisers may notice they're getting very little traction from your site. Page views go up, but uniques go down.

Have you met Reddit? They love flamewars.

I thought they were having advertiser issues, but maybe I'm out of the loop here.

Re: Deriving the Reddit Formula

#15

Post by keysersosa » Wed, Jul 15, 2015, 5:22 AM UTC

tim333 wrote: post #10

>This resolves one of the original mysteries — why the current time doesn't appear in Reddit's formula. The current time did appear in the very early Reddit's formula but they figured it was computationally much faster to inflate the scores for new stories than to have to go through the database reducing the scores for loads of old stories.

Confirmed, but the time dependent model didn't ask for long. I seem to recall switching to the current model (with slightly different constants) during the switch to python in late '05.

That said, one of the early "mistakes" we made was not noticing the obvious step of taking the log of on the exponential form formula, which ended up meaning we had only a couple hundred days before we started to run afoul of the max float size in postgresql. I also seem to recall coming up with that obvious trick with only a few days to spare.

Re: Deriving the Reddit Formula

#16

Post by jedberg » Wed, Jul 15, 2015, 5:28 AM UTC

keysersosa wrote: post #15

tim333 wrote: post #10

>This resolves one of the original mysteries — why the current time doesn't appear in Reddit's formula. The current time did appear in the very early Reddit's formula but they figured it was computationally much faster to inflate the scores for new stories than to have to go through the database reducing the scores for loads of old stories.

Confirmed, but the time dependent model didn't ask for long. I seem to recall switching to the current model (with slightly different constants) during the switch to python in late '05. That said, one of the early "mistakes" we made was not noticing the obvious step of taking the log of on the exponential form formula, which ended up meaning we had only a couple hundred days before we started to run afoul of the max…

I'm really glad you're here, because I was about to comment that you're the only person I know who understands this stuff.

(for those that don't know keysersosa was the guy who wrote the current hot algorithm)

Re: Deriving the Reddit Formula

#17

Post by studentrob » Wed, Jul 15, 2015, 6:46 AM UTC

Cool. But why is an outside researcher writing about this? Where is the research from within Reddit? They have as unique a dataset as Facebook, google or yahoo. They should do something cool with it and share it, like this guy

Re: Deriving the Reddit Formula

#18

Post by nostrademons » Wed, Jul 15, 2015, 8:02 AM UTC

gohrt wrote: post #12

tedunangst wrote: post #7

Earlier quoted context omitted.

Possibly short term, but if you build a site around the kinds of people who like flame wars, your advertisers may notice they're getting very little traction from your site. Page views go up, but uniques go down.

Have you met Reddit? They love flamewars.

Yishan posted a couple hours ago that Reddit's business model is built upon flamewars. Basically get people heated up and then they gild the comments they agree with.

https://np.reddit.com/r/announcements/comments/3dautm/conten...

...come to think of it, this explains a lot of the Reddit drama. The Reddit Gold was flying left and right during the black-out (despite pleas not to give any money to the site), and then every time an admin, CEO, former CEO, or board member pours gas on the fire, their revenue goes up. It's brilliant! They've figured out how to make money off of hurt feelings and angry people, and now have an incentive to cause as much drama as possible.

Re: Deriving the Reddit Formula

#19

Post by RickHull » Wed, Jul 15, 2015, 8:18 AM UTC

nostrademons wrote: post #18

gohrt wrote: post #12

Earlier quoted context omitted.

Have you met Reddit? They love flamewars.

Yishan posted a couple hours ago that Reddit's business model is built upon flamewars. Basically get people heated up and then they gild the comments they agree with. https://np.reddit.com/r/announcements/comments/3dautm/conten... ...come to think of it, this explains a lot of the Reddit drama. The Reddit Gold was flying left and right during the black-out (despite pleas not to give any money to the site), and then e…

Ha, this is great, in a not-so-great but kinda-great way!

Re: Deriving the Reddit Formula

#20

Post by tim333 » Wed, Jul 15, 2015, 11:28 AM UTC

keysersosa wrote: post #15

tim333 wrote: post #10

>This resolves one of the original mysteries — why the current time doesn't appear in Reddit's formula. The current time did appear in the very early Reddit's formula but they figured it was computationally much faster to inflate the scores for new stories than to have to go through the database reducing the scores for loads of old stories.

Confirmed, but the time dependent model didn't ask for long. I seem to recall switching to the current model (with slightly different constants) during the switch to python in late '05. That said, one of the early "mistakes" we made was not noticing the obvious step of taking the log of on the exponential form formula, which ended up meaning we had only a couple hundred days before we started to run afoul of the max…

Ah - I was thinking about how exponential growth would go over the max size and how to deal with it and also wondering why they were taking logs and not putting two and two together. It's surprising how tricky it can be to see what seems obvious with hindsight.