Assassinator Wrote:EDIT: read a bit more, seems like they want to use Pearson's correlation and scale the output from (-1,1) » (0,1).

So yeah, in MyAnimeList, a 0% would mean perfect negative correlation, 50% = no correlation, 100% = perfect correlation.

Now it's almost impossible to get much negative correlation at all, because that would mean you hate everything the other person loves, vice versa (which is extremely unlikely, because everyone likes the good anime, and hates the bad anime, and that = positive correlation).  So you can say it pretty much starts at 50%.

So my 62% with Sensei and Sky aren't really high at all (it's only relatively mild correlation).  However, my 91% with my sister IS pretty damn high.

Heartless141 Wrote:sigh, i can't even add em all to my list, -.- you tell me to go rate em?

Yes?

I want to see my compatibility with you, lol.

(This post was last modified: 02/11/2009 06:01 AM by Assassinator.)

Assassinator Wrote:So my 62% with Sensei and Sky aren't really high at all.  However, my 91% with my sister IS pretty damn high.

Medium-High is what it spits out though.

I'll be completely honest, I thought it was gonna be much lower.. that's sort of nice to know to be honest =^-^=

Senseito7 Wrote:

Assassinator Wrote:So my 62% with Sensei and Sky aren't really high at all.  However, my 91% with my sister IS pretty damn high.

Medium-High is what it spits out though.


I'll be completely honest, I thought it was gonna be much lower.. that's sort of nice to know to be honest =^-^=

"Medium-High" is a lie, lol.

According to the maths and the data (anime), it should be more like

0-30% - Doesn't exist. it's impossible to get this
30-40% - extremely incompatible
40-50% - incompatible
50-60% - not very compatible
60-70% - average compatibility
70-80% - good compatability
80-90% - extremely compatible
90-100% - ultra compatible.

So yeah, it is actually much lower...

guy who implemented the algorithm Wrote:Ah, I forgot to mention. Wee took the original score from -100% to 100% and divided it by 2, then added 50. So, 28/2 = 14+50 = 64%. Apparently still not the same as the 56% detailed, but it's closer. This method allows the users to always see a positive percentage.

Wee're actually only 24.4% according to the original scale.

(This post was last modified: 02/11/2009 06:49 AM by Assassinator.)

Assassinator Wrote:Wee're actually only 24.4% according to the original scale.

Don't try to make me sad ^^'

(This post was last modified: 02/11/2009 06:24 AM by ProperBritish.)

So Pearson correlation test with scaling applied?  Not really that effective.  Perhaps would be better if slightly scaled with anime similarities.
Oh well, I guess people like random stats anyway.

My star system doesn't really scale to an out of 10 scale...  It's aimed at a fair distribution, but in reality, you'd never rate an anime less than 3/10 unless you love torturing yourself.

(This post was last modified: 02/11/2009 06:27 AM by ZiNgA BuRgA.)

Senseito7 Wrote:

Assassinator Wrote:Wee're actually only 24.4% according to the original scale.

Don't try to make me sad ^^'

Why would that make you sad?

---

Correlation examples... Look at the top line.

me,you = 0.24.
you,sky = 0.64
me,sister = 0.82

ZiNgA BuRgA Wrote:So Pearson correlation test with scaling applied?  Not really that effective.  Perhaps would be better if slightly scaled with anime similarities.
Oh well, I guess people like random stats anyway.

Which is what I'm saying.  You can't really (well, extremely difficult to) get negative correlations with anime, in MyAnimeList sense, translates to you can't really get below 50%, which imo is pretty stupid.

What needs to be done is that instead of each anime being scaled according your distribution, it needs to be scaled with the distribution of all the other scores for that particular anime (to get rid of "good anime gets good scores form everyone").  This way it should give better correlation figures.... if stuff doesn't get warped too much.

(This post was last modified: 02/11/2009 06:55 PM by Assassinator.)

Assassinator Wrote:What needs to be done is that instead of each anime being scaled according your distribution, it needs to be first scaled with the distribution of all the other scores for that particular anime (to get rid of "good anime gets good scores form everyone"), then scale that result with your own distribution.  This way it should give better correlation figures.

Why not suggest it?

Senseito7 Wrote:

Assassinator Wrote:What needs to be done is that instead of each anime being scaled according your distribution, it needs to be first scaled with the distribution of all the other scores for that particular anime (to get rid of "good anime gets good scores form everyone"), then scale that result with your own distribution.  This way it should give better correlation figures.

Why not suggest it?

Because that's all talk and no proof.  I'd need to study up on a whole crapload of statistical theory, to make sure everything is all right, and then provide them with some solid maths.  And I really can't be stuffed doing that.

Oh, and that also might be quite computationally intensive.  I mean some anime have like 50000 votes on it, and lets say you have 100 anime in your list that matches... sounds heavy.  Although since I'm not much of a programmer, I can't really say for certain (ask Zinga), but it's probably not that viable.

(This post was last modified: 02/11/2009 07:18 AM by Assassinator.)