What is pouets logic about the #rank in the alltime-top?
category: general [glöplog]
@Gargaj(?):
What is pouets logic about the #rank in the alltime-top? Only the number of comments?
And how often is this number updated?
What is pouets logic about the #rank in the alltime-top? Only the number of comments?
And how often is this number updated?
Quote:
And how often is this number updated?
Once a day, every 5AM.
Code:
// this needs to be made faster. a LOT faster.
Do it in a single SQL statement, using a window function?
If I understand this right, the logic is ranking by rank(views)+rank(votes), but I don't honestly know what a view is here. Every time someone looks at a prod page, it increments a counter in the database?
Yes. And yes it should be a single SQL statement, but priority was getting Pouet2 off the ground rather than how fast a crontab runs. 😅
If we wanna change the system btw, I'm not against it - my experiments with a Wilson Score system are interesting at least - but I'm not hopeful that there'll be any sort of consensus on how to change something that was set in stone for 20 years.
If we wanna change the system btw, I'm not against it - my experiments with a Wilson Score system are interesting at least - but I'm not hopeful that there'll be any sort of consensus on how to change something that was set in stone for 20 years.
How can you even have a Wilson score interval of a trinomial? (FWIW, it seems the implementation is missing continuity correction; it only lines up with R's prop.test() if I set correct=FALSE)
FWIW; I did a quick test locally based on the data I have (the JSON dump, which doesn't have views), and it updated all the ranks in about seven seconds. I guess that ought to be fast enough.
FWIW; I did a quick test locally based on the data I have (the JSON dump, which doesn't have views), and it updated all the ranks in about seven seconds. I guess that ought to be fast enough.
Code:
pouet=# with t1 as (
select id, rank() over (order by ((data->>'voteup')::integer - (data->>'votedown')::integer) desc) as new_rank from prod_test
)
update prod_test set my_rank=t1.new_rank from t1 where prod_test.id=t1.id;
UPDATE 92757
Time: 6758,497 ms (00:06,758)
Quote:
How can you even have a Wilson score interval of a trinomial?
I mean the formula considers "number of positive votes" and "number of total votes" so you can get there in all sorts of ways (e.g. adding neutral and negative votes together, or subtracting negative votes from the positive count, etc etc)
Would be very interesting to have a list of all pouet-prods (or the top1000) ranked only by positive votes.
If I did it correctly and according to all the positive votes of the top 25 this would be the news top 10. Not so much change. Only »1995« ranked up from #13 and entered the top 10 while »tube« left it.
01 (01) – 811 – Farbrausch – Debris
02 (04) – 657 – ASD – Lifeforce
03 (03) – 635 – Exceed – Heaven Seven
04 (02) – 614 – Rgba & TBC – Elevated
05 (08) – 568 – The Black Lotus – Starstruck
06 (05) – 559 – Farbrausch – The Popular Demo
07 (07) – 543 – Orb & Andromeda – Stargazer
08 (13) – 532 – Kewlers & MFX – 1995
09 (09) – 524 – Equinox – Kings Of The Playground
10 (06) – 506 – Farbrausch – The Product
If I did it correctly and according to all the positive votes of the top 25 this would be the news top 10. Not so much change. Only »1995« ranked up from #13 and entered the top 10 while »tube« left it.
01 (01) – 811 – Farbrausch – Debris
02 (04) – 657 – ASD – Lifeforce
03 (03) – 635 – Exceed – Heaven Seven
04 (02) – 614 – Rgba & TBC – Elevated
05 (08) – 568 – The Black Lotus – Starstruck
06 (05) – 559 – Farbrausch – The Popular Demo
07 (07) – 543 – Orb & Andromeda – Stargazer
08 (13) – 532 – Kewlers & MFX – 1995
09 (09) – 524 – Equinox – Kings Of The Playground
10 (06) – 506 – Farbrausch – The Product
Quote:
but I'm not hopeful that there'll be any sort of consensus on how to change something that was set in stone for 20 years.
But perhaps you could add some checkboxes/pulldowns at the »Top of the trumpets«-page, so it's possible to see the list in different ways.
I think the algorithm to select these is broken and/or too simple. Being always visible on the front page gave these prods too much space in the spotlight, making them artificially popular and now the algorithm keeps picking them forever. It has been practically a static list for years.
It's like training a neural network with data generated by the network itself.
Maybe an idea worth trying would be to just add noise to the ranking, so that atleast visibility would distribute a little bit more meaningfully to the best prods of Pouet? Although the biasing problem this kind of frontpage all-time top-list causes, might be fundamentally unfixable.
It's like training a neural network with data generated by the network itself.
Maybe an idea worth trying would be to just add noise to the ranking, so that atleast visibility would distribute a little bit more meaningfully to the best prods of Pouet? Although the biasing problem this kind of frontpage all-time top-list causes, might be fundamentally unfixable.
Part of the problem is that the site's peak popularity set some impressive numbers on the vote counts that'll be hard to repro without the same amount of active users.
Peak popularity and voting seems to be around mid-late 00's. You could make a rotating turd with noise and still get 200 thumb ups. I don't think many windows demos got to 200 thumb ups in the last 8 years alone (* only two).
At least grouping by added date, this seems to be very true:
Of course, in this formulation, a 2006 prod could get a 2023 vote and it would count.
Just being on the front changes voting patterns, though (exposure to different groups of voters). There are lots of prods that I find “meh” to the point of not even bothering to leave a piggie, but I would probably have commented on them if they were on the front page (moving the average vote down). The endgame would, of course, be personalized recommendations by means of AI, which everybody loves to hate. :-)
Code:
year | total_votes
------+-------------
2000 | 34009
2001 | 52497
2002 | 50530
2003 | 45800
2004 | 51118
2005 | 59490
2006 | 65355
2007 | 60654
2008 | 58400
2009 | 44157
2010 | 40016
2011 | 30329
2012 | 33605
2013 | 27775
2014 | 30186
2015 | 27497
2016 | 22897
2017 | 25016
2018 | 24308
2019 | 24518
2020 | 18396
2021 | 18656
2022 | 17564
2023 | 11555
(24 rows)
Of course, in this formulation, a 2006 prod could get a 2023 vote and it would count.
Just being on the front changes voting patterns, though (exposure to different groups of voters). There are lots of prods that I find “meh” to the point of not even bothering to leave a piggie, but I would probably have commented on them if they were on the front page (moving the average vote down). The endgame would, of course, be personalized recommendations by means of AI, which everybody loves to hate. :-)
Quote:
I don't think many windows demos got to 200 thumb ups in the last 8 years alone (* only two).
Well, if you count intros, there are five more.
Only 59 demos ever have 200+ thumbs, so turd-with-noise may be a slight exaggeration, but there certainly is some truth to it.
Yikes, seems like COVID killed the scene :|
[qoute]
2019 | 24518
2020 | 18396
[/qoute]
[qoute]
2019 | 24518
2020 | 18396
[/qoute]
Maybe a good way to make the all time-top less static would be to add simple karma decay akin to reddit.com, and this could also work for top of the month, something like that:
- For all time-top, prods lose one point per month in the ranking formula.
- For top of the month, prods lose one point per day in the ranking formula, that way you don't even need a limit to 30 days, could also be renamed 'top of the moment'.
My two cents :)
- For all time-top, prods lose one point per month in the ranking formula.
- For top of the month, prods lose one point per day in the ranking formula, that way you don't even need a limit to 30 days, could also be renamed 'top of the moment'.
My two cents :)
By the way ... the reason I started this thread was that I didn't understood, why some prods are ranked higher than others. Oh, and it's not that I really care where my or other prods are ranked. It's just out of curiosity how the system works.
For example:
Focused by Gaspode: 53 upvotes, 4 piggies, 0 downvotes –> #5133
Aquaverse by Rebels & Calodox: 39 upvotes, 6 piggies, 0 downvotes –> #3805
This is a difference of 1328, while my prod has 14 more upvotes.
For example:
Focused by Gaspode: 53 upvotes, 4 piggies, 0 downvotes –> #5133
Aquaverse by Rebels & Calodox: 39 upvotes, 6 piggies, 0 downvotes –> #3805
This is a difference of 1328, while my prod has 14 more upvotes.
you may want to keep the traditionally "top" prods (ie debris) at front page of pouet as the offering of the "best" of the scene, maybe ever.
Recent demos may just be recent but not necessarily best
Recent demos may just be recent but not necessarily best
gaspode: I'm not sure if you are trying to say you understood now or if you still don't know, but (rank of) number of views also seemingly factor in. :-) And you cannot see those stats except indirectly.
@Sesse: No, I still don't understand.
I understand that the number of pageviews is a factor. But how could it be such a big factor? We are talking about a difference of 1328 steps between two prods that were released in the same compo.
I understand that the number of pageviews is a factor. But how could it be such a big factor? We are talking about a difference of 1328 steps between two prods that were released in the same compo.
From what I can see, based on the latest data I have available (won't be exactly in sync with live Pouët):
- Focused is at vote-rank #2294
- Aquaverse is at vote-rank #3441
- Focused is at viewer-rank #10699
- Aquaverse is at viewer-rank #5655
- Total score for Focused: 12993 (lower is better)
- Total score for Aquaverse: 16354
This is assuming “popularity” is monotonic with number of viewers. So Aquaverse is far “better” in total score because of the large discrepancy in viewer count, so it's not surprising it's ranking a lot higher.
- Focused is at vote-rank #2294
- Aquaverse is at vote-rank #3441
- Focused is at viewer-rank #10699
- Aquaverse is at viewer-rank #5655
- Total score for Focused: 12993 (lower is better)
- Total score for Aquaverse: 16354
This is assuming “popularity” is monotonic with number of viewers. So Aquaverse is far “better” in total score because of the large discrepancy in viewer count, so it's not surprising it's ranking a lot higher.
Quote:
Focused is at viewer-rank #10699
Does this mean I get »punisehd« for checking if i got comments? (aka views)
Although I don't think I checked for 10k times. :o)
Lower rank is better, so no. #1 in viewer-rank (highest in popularity) is Debris. And yes, I checked that this number comes directly from number of viewers, so the rank should be the same even though we don't get Pouët's absolute viewer counts.
I see that I completely messed up the last sum there. Total score for Aquaverse is 9096.
But why is it better that the number of views is lower? aka less people visiting the prod-page.