Anime Hints Mk. II

nstgc · Post by **nstgc** » Sun May 22, 2005 6:51 am

Firstly, sorry about the cheezy title for this thread (it was that or shin anime hints).
On the subject, I finaly got around to looking at how the anime hint system works. I am dissatisfied. While I am complaining, I have a solution.

I have already worked out all the math and am wondering if anyone is interested and if so how do I put my formula into action.

The two main changes are how user weight is calculated and tweaks to scoring. I believe these lead to a more accurate, and faster (simpler script) method.

Post by **exp** » Sun May 22, 2005 9:52 pm

egg is doing the anime hint part.
so you should direct your suggestions at him.
the best way would probably to write an extensive forum post here and sending him a forum pm with the link to the topic.

BYe!
EXP

nstgc · Post by **nstgc** » Mon May 23, 2005 4:38 am

could i attach a pdf file.

also, (i really hate asking this), not to be insulting, but what is egg's math proficiency? People tell me all the time that I assume they know more then they actualy do. anyway, I'll write it up in word. If I can post it as a pdf file I will, other wise i'll see if I can write it as a methematica "program" (I guess you can call it that). I know it can't run on a server, but using words alone to convey math is difficult I think. I'm just overjoyed someone listened to me for once.

One last thing; who is egg and how do i contact him?

Andemon · Post by **Andemon** » Mon May 23, 2005 4:50 am

Which reminds me -- I've heard that Anime Hint will be revamped after the new category system is finished. Not sure if that's true or not, but if it is, I suppose it'd be somewhat useless to base your changes on the current system...

...and exp already told you how to contact egg -- pm (private message) him.

(I wouldn't mind if you could increase the accuracy, but frankly, the current version of Anime Hint doesn't seem as bad as you make it sound. The recommendations that I get from the it seem to be accurate enough.

It doesn't support *.5 votes, though -- but I suppose egg has his reasons for not adding that yet...)

nstgc · Post by **nstgc** » Mon May 23, 2005 5:25 am

Unless the wiki is wrong, it sounds very logic based and statistics should be done with minimal logic, even if its computer driven.

also, i don't know who egg. sorry. i'll check since i just finished my final draft of the word document (meaning all thats left is the pdf conversion).

Post by **exp** » Mon May 23, 2005 7:14 am

yes the new category system will make more data available to be used in calculating anime recommendations.
however the main part of the calculation will not change. there will most likely only be some additional parts for the calculation.
so it's no reason to stop working on the anime hint feature atm.

to reach egg use:
http://www.anidb.net/forum/privmsg.php?mode=post&u=648

BYe!
EXP

egg · Post by **egg** » Mon May 23, 2005 5:43 pm

Andemon wrote:Which reminds me -- I've heard that Anime Hint will be revamped after the new category system is finished. Not sure if that's true or not, but if it is, I suppose it'd be somewhat useless to base your changes on the current system...

Where did you hear that??? I have no idea how to incorporate the categories in the hint... I have just been kinda busy, and haven't worked on the hint for a while. I keep meaning to do it though, I will probably get to it in the next few weeks. What I expect to be there is:

Support for .5 voting for existing hint calculation logic
A new more mathematically/statistically sound method for comparing users
A filter for animes with fewer than x users to vote
Maybe more depending on how ambitious I get

Andemon · Post by **Andemon** » Mon May 23, 2005 6:09 pm

egg wrote:
Andemon wrote:Which reminds me -- I've heard that Anime Hint will be revamped after the new category system is finished.
Where did you hear that??? I have no idea how to incorporate the categories in the hint...

Well, er... Perhaps you'd better ask exp about it. -_-;
...but as you can see from his above post, I misunderstood the situation somewhat. It's just that he has mentioned once or twice that one of the major benefits of the new category system will be that it'll be useful in making recommendations (...or something like that). Sounded to me like it'll mean major modifications for Anime Hint.

I can see that there's potential for improving the hint with the new category system, though. For example, if the Dynamic::Plot::Epic and Dynamic::Plot::Episodic get added, then wouldn't it be possible to determine from those and the users votes which type of plots the user prefers -- then slant the recommendations based on that...?

egg · Post by **egg** » Mon May 23, 2005 7:32 pm

Andemon wrote:then slant the recommendations based on that...?

How? I need equations or code to correlate those. I can build a table of User preferences using a formula like:
score for a category = sum for all animes of (category weight * user vote) / (sum for all animes (category weights) * num of animes)

Here the category weight is the 1-3 stars used to determine the significance of that category. Then how do I incorporate these scores into the hint?

Andemon · Post by **Andemon** » Tue May 24, 2005 3:22 am

egg wrote:Here the category weight is the 1-3 stars used to determine the significance of that category. Then how do I incorporate these scores into the hint?

So let's say that it has been determined that the user likes ::Epic anime -- you could always generate the recommendations normally first, then check that list for animes that are ::Epic, and give them bonus points based on their star-rating so they'll appear higher up on the list... As to how to actually do that and with what weights -- I have no idea about the inner workings of Anime Hint, so I can't comment on that.

Of course, it wouldn't be advisable to use every category in such way (fe. Characters::Animals::Hamsters -- just because someone has voted high for Hamtaro, doesn't mean he'll like Oruchuban Ebichu... -_-), but Dynamic:: seems well-suited for it...

egg · Post by **egg** » Tue May 24, 2005 6:29 am

Whatever system is created, it will NOT be category knowledgable. I will not make it to look at certain categories and not others (with the possible exception of Genre::Hentai. There will probably hundreds of rated categories per user (after a user has voted on 50 or so anime that WILL be a lot of categories). My guess is that animes on average will have about 10 categories. The question is how to adjust a rating based on the user's ratings of the categories for all of the categories for the new anime. Although the system will not pay attention to specific categories, my guess that based on other attributes (similar users, ...) and the OTHER categories in the animes, just because someone gave Hamtaro a 10 doesn't necessarily mean that Oruchuban Ebichu will be recommended.

nstgc · Post by **nstgc** » Wed Jun 15, 2005 1:30 am

While i wait for egg's reply (i messaged him several days ago) i would like other people's oppinion of my work. its been finished for some time now and i am eager to get feed back.

unfortunely, due to my poor righting skills, i ask that those posting their ideas be certian that they understood what i wrote. this means have a decent amount of knowlegde and alot of creativity (other wise known as true intelegence). creative thinking is, by far, more important then mere knowlegde.

take the difference between engineers and physicists. engeneers are ment to remember information and do, in physics we mainly must think. both, however, have their places. my brother's place is as an engineer making loads of dough, mine to to think so abstractly my thoughts will not be useful for another 50-75 years. unless i'm directing my research towards solid-state physics which does have immediate uses.

if you cannot veiw the document and you have the proper software please tell me!

the url is [url]http://nst_gc.tripod.com/Ani-Hints_Mk_III.pdf[/url]
it was working the other day, but its not working for me right now, but my computer is not doing well at all so who k
nows.

egg · Post by **egg** » Wed Jun 15, 2005 2:55 am

Well, I read through it. To tell you the truth, I was a bit lost, but the example helped me get the gist of it. It would still take me a while to go through and figure out the details though. I think that I need to concentrate on getting out the current hint (yes with the correlation coefficients because that is already coded). After a while I will come back and look at this again. The next step for me would be to try to do a mock up in excel (I think it can handle everything since you have the approximation functions). That way I can compare it to existing methods and get an idea of the calculations necessary and the load it will put on the server.

If you wanted to do that for me, I would probably take a look at it sooner. Unfortunately this is drastically different from the other ones and will require a rewrite of some section of code to handle it, so unless I can see what appears to be an improvement in the results from some sort of mock up, I probably will not put in the effort to code it. I am keeping an open mind, but I also want to be honest with you and don't want you to get your hopes up in case I [and exp] decide not to do it.

I hope other people will also comment so I can get some other input...

nstgc · Post by **nstgc** » Wed Jun 15, 2005 3:33 am

so the corrilation coeffecient method has already been coded?

egg · Post by **egg** » Wed Jun 15, 2005 4:44 am

nstgc wrote:so the corrilation coeffecient method has already been coded?

Yes, it was most of the way done when you started proposing your solution. Currently it is on my dev machine, I am also working on a few other things (support for .5 votes, a few more filters, ...). Once I am finished with that I will send it to exp for some testing and then exp will put it online when he is ready. After that I will look at your solution again.