if i had ranked elo data and tag metadata for 20,000 anime images

File: __assassin_of_black_and_illyasviel_von_einzbern_fate_apocrypha_fate_kaleid_liner_prisma_illya_and_fate_series_drawn_by_wakaba_wkbast__e00e6411868238c31f11d684585cf1c5.jpg (117KB, 717x1024px) Image search: [Google]

117KB, 717x1024px

have you tried updoc?

File: Frame2279.jpg (187KB, 1920x1080px) Image search: [Google]

187KB, 1920x1080px

>>60656695
whats updoc

>>60656720
>w<

>>60656583
> elo data
> images
Seems shit, get TrueSkill data eh.

And what statistics do you need if the images are already rated and tagged? You gonna try and find out if people like a tag not so dependent on how good an image is or what?

>>60656859
the elo rating is all me; so the data can be used to help find cute images on boorus and stuff

>>60656877
Ya, use TrueSkill or such, not shitty ELO. Even in muh US software patent land it should be okay cause MS seems to allow noncommercial use and you can just plug an open sauce library.

That said I don't even know how you want to get to cute rather than figuring out tag combination etc preferences which are kinda trivial. Are you trying to wrangle this through a human emotion /perception model that does not yet exist in this complexity and then tweak the parameters of it until it is how (You) personally prefer it with AI algorithms?!

>>60656950
i already have 20,000 training data of how cute the image is and all the tags; surely all i need to do is smash them together and if i get the fitness or whatever high enough and i dont overfit then magically i can help find cute images via tags right

what am i missing here
and truskill looks interesting; i can rewrite my thingy to use it in the future

>>60656950
PS: Why Trueskill? Because you have a FFA type of situation, not 1v1. It generally converges on a sane rating much faster and all that.

>>60656981
You intend to run your fitness functions on what?

Yea can extract which tags you like best sure, but what are the odds that you don't already know...?

I don't know what else you'd do, its not like you currently have a model that does simulate your perception adequately in such a way that you might be able to train and use it to score new images.

>>60657122
the cute images were calculated entirely through 1v1s; in sequence like a tournament with stronger and stronger images fighting them until they lose. Or sometimes random comparisons; and sometimes other systems. So that particular feature of trueskill doesn't help

>>60657122
>Yea can extract which tags you like best sure, but what are the odds that you don't already know...?

1. I want an automated system; so it's either REAL statistics which i dont want to do or just cheat

fitness function on prediction of how cute the image is based on tags

could also find combinations of tags that make things cute and other interesting things

sorry ive never made an ANN before just thought about it alot

>>60657122
By the way ACQUINE will not be good and the photofeeler type of approach is next to pointless since even if you had a superbly trained auto tagging neural network it'll only mildly augment the huge more accurate sauce that are the boorus.

>>60657191
I don't intent to try and identify tags just correlate a list of all tags with the final cuteness rating

so if I get a list of tag from booru it can vet the images before they reach my eyes

>>60657157
> 1v1
You will need to do the 1v1 thing with ELO for a long time with many games until you have somewhat accurate data. Unless you got a bunch of really simple preferences I'd really guess you still rated very many too high or low.

> fitness function on prediction of how cute the image is based on tags
That will result in photofeeler ridiculousness because it really has no sense of aesthetics and indeed its not really done with a fitness function, you just deduce the appropriate score weight for each of the current tags and when you see a new image you can get the sum of scores for that image.

> could also find combinations of tags that make things cute and other interesting things
You could Trueskill for the combinations.

>>60657327
ELO was higher than 95% presictive at the end; I did lots of comparisons

Hundreds of thousands
Using trueskill and automatically smashing teams of images together is a really smart idea a Anon Thanks

I'm going to sleep now

But that method ^ would need an infinite number of teams to find combos; it could just rate individual tags

>>60657379
Of course full optimization of tag score groups even when you already have perfect scoring will still be a NP-Hard "can't really be done" knapsack type of problem. You can test interesting candidates though. Are the best pairings of tags combined as good as you'd think, is there any point to the exact ordering of some of the tags or might they as well be a group with exactly the same score etc.

Anyhow good night!

>>60656720
Nm u