Regarding the recent site issues (502s) and Bayesian averages and ratings histogram

Status
Not open for further replies.
Fed-Kun's army
Joined
Feb 9, 2018
Messages
326
@definitionofinsanity @floemizu unless you are running a local proxy, 502 never means that your connection is broken. It's a response from the gateway (in this case cloudflare) that the backend service (mangadex) did not respond or made an error. ?
 
Aggregator gang
Joined
Jan 21, 2018
Messages
270
One issue with the Bayesian average is that with a straightforward implementation it will end up pulling really bad manga too close to the mean. It wouldn't be a problem normally but in rating you want bad stuff at the bottom, not near the middle. In systems where the real average isn't the expected average (real average is actually almost 8 but the implicit expected average is 5.5) low ratings get pulled too high, when ideally predominantly low ratings would be weighted more.

This does work if you want to encourage people to read the low manga more and get it to its actual rating more quickly, but I think it's bad user experience, especially paired with the seeming lack of relation between the histogram and the final rating.

I don't know where that fudge factor would end up or to what manga it would apply, but it might be best to increase the weight of low ratings for manga with lower numbers of reviews. The problem with this is that a manga with only 1 star ratings would have their rating increased with each new 1 star review. Maybe scaling the distribution from the real Bayesian to the expected average before calculating would work and then mapping back, but maybe it would do nothing at all. I might do some MATLAB to figure it out later.

It may be intended, but sorting manga by rating is still based on the mean rather than the Bayesian average.

EDIT: The mapping trick does not work. As I feared, it doesn't do anything in the end.
 
Miku best girl
Admin
Joined
May 29, 2012
Messages
1,441
I'll update the rest of the site once we agree on the Bayesian formula.
 
Dex-chan lover
Joined
Jun 22, 2018
Messages
2,127
@SuppeBargeld
I wasn't sure because since upgrading to Windows 10 (unfortunately and not by choice) I've been getting random and infuriatingly often disconnects with internet connection troubleshooter being the easiest way to fix it and citing "the default gateway not available" being the issues and resetting the wifi. Never had that issue before until Win 10.
 
Contributor
Joined
Jan 18, 2018
Messages
2,156
Bayesian, That's I just learned exactly 1.5 years ago in second semester.
and I forgot it all ?
 
Dex-chan lover
Joined
Jan 22, 2018
Messages
1,419
maam/sir, question...

based on the calculation above changing manga instead of restaurant... the "C" in the calculation means all manga?...
like if "R" is average rating of, let say One Piece, then "C" is the average rating of all the manga titles in mangadex?...
 
Aggregator gang
Joined
Jan 21, 2018
Messages
270
So doing a Bayesian average with just Kaguya and a hypothetical terrible manga with only 500 1 star ratings still gives the terrible manga a 2.8 rating.

The best way to do it, I think, is to fudge w a little bit. It's essentially a measurement of how much you allow the character of the individual versus the conglomerate to show through. The values are normally [0,1.0] but if you range it from [0.5,1.0] you guarantee half of the rating is due to the individual manga. There's still the problem that even my hypothetically detested manga ends up with a 2 star rating unless you make the floor so high that a low number of ratings can tank the average.

To fix that, I tried a variable weighting. If the number of ratings for one manga is less than the average number of ratings (m), then there's a scaling weight from 0.5 to 0.85. If it's equal or more, it's just a flat 0.85.

PXpJIpw.png


@zetsuravez
C is the average rating overall. It's like if you take the ratings from every single manga and combine them into one big bunch and take the average.
 
Joined
Aug 8, 2018
Messages
185
@SuppeBargeld , agreed the poor excuse of traffic is crap as its the use of cloudflare which is a pro hacker server which allows stressers, ddos's, and other frequent use of their services which causes websites to have trouble, this moving to a bigger server but still using cloudflare isnt helpful, it would do well if more intelligent people understood this, their may be reasons to use cloudflare but blaming traffic and not the hack friendly servers isnt right either
 
Power Uploader
Joined
Jan 26, 2018
Messages
103
I think the issue of skewed ratings will never be adequately addressed. A purely open rating system is fine itself but what's needed is a group of dedicated critiques' that are willing to sit down and page through series and provide an opinion. Similar to Metacritic/Rotten tomatoes with a public pool and trusted pool. Problem is that there really aren't any sites for this and even then trusting a cherry picked group of these opinions is kinda a shot in the dark, as there are plenty of false critique score-bloat reviews on both examples above.

As for the current system an algorithm for assessment will always be flawed as people find ways to manipulate them and even assuming it's a "correct" numeric score the majority of people will disagree with it. Additionally I do think it's foolish using a mean or adjusted mean as Bayasian given it's known statistical issues. Ironically the best solution to finding an appropriate score is to not bother. Simply provide a data chart; as Holo has done, and let people form their own judgement.
*An example being my tendency to trust the median and ignore scores of 1 or 10 in the majority of cases.
 
Dex-chan lover
Joined
Jun 28, 2018
Messages
526
Happy New Year! Is it possible to search for a specific genre? I mean without any keywords, just select the genre you want, then hit search and the engine will spit out the mangas/manhuas of the relevant genre. If not it would be great if you could make it happen as right now when you don't know what title you want to search only it's genre the engine won't spit out anything, you have to give a keyword from what I saw.
 
Contributor
Joined
Mar 10, 2018
Messages
3,499
Thanks for your efford Holo and end Doki team. Just donate some $ for this new year. ?
 
Miku best girl
Admin
Joined
May 29, 2012
Messages
1,441
There will always be problems with trying to be objective about ratings, which is subjective... There will always be someone who disagrees.

Even ratings are inherently biased towards the positive end because you're more likely to rate titles in genres that you like.
 
Dex-chan lover
Joined
Feb 5, 2018
Messages
791
@Holo
Have you tried setting up page rules? I heard Clouldflare doesn't cache HTML content by default. Maybe it could help soak up some incoming load.

DGaCG7M.png
2t34WIP.png
 
Miku best girl
Admin
Joined
May 29, 2012
Messages
1,441
Cloudflare already caches 75% of our traffic. The html won't make of a dent.
 
Dex-chan lover
Joined
Feb 5, 2018
Messages
791
@Holo
You wouldn't know unless you try. A 5-8% increase in caching might not be bad.
 
Dex-chan lover
Joined
Sep 3, 2018
Messages
1,071
Thanks for you're hard work, and I hope you guys remove the 20 page preloading limit and make the option of loading the whole chapter at once available
 
Status
Not open for further replies.

Users who are viewing this thread

Top